1 | namespace dnnl { |
2 | namespace impl { |
3 | namespace gpu { |
4 | namespace ocl { |
5 | const char * = R"==(/******************************************************************************* )==" "\n" |
6 | R"==(* Copyright 2019-2022 Intel Corporation )==" "\n" |
7 | R"==(* )==" "\n" |
8 | R"==(* Licensed under the Apache License, Version 2.0 (the "License"); )==" "\n" |
9 | R"==(* you may not use this file except in compliance with the License. )==" "\n" |
10 | R"==(* You may obtain a copy of the License at )==" "\n" |
11 | R"==(* )==" "\n" |
12 | R"==(* http: )==" "\n" |
13 | R"==(* )==" "\n" |
14 | R"==(* Unless required by applicable law or agreed to in writing, software )==" "\n" |
15 | R"==(* distributed under the License is distributed on an "AS IS" BASIS, )==" "\n" |
16 | R"==(* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. )==" "\n" |
17 | R"==(* See the License for the specific language governing permissions and )==" "\n" |
18 | R"==(* limitations under the License. )==" "\n" |
19 | R"==(*******************************************************************************/ )==" "\n" |
20 | R"==(#ifndef GPU_OCL_OCL_TYPES_H )==" "\n" |
21 | R"==(#define GPU_OCL_OCL_TYPES_H )==" "\n" |
22 | R"==(#include "gpu/ocl/ocl_math_utils.h" )==" "\n" |
23 | R"==(#define unroll_for __attribute__((opencl_unroll_hint)) for )==" "\n" |
24 | R"==(#define for_ for )==" "\n" |
25 | R"==(#define CONCAt2(a, b) a##b )==" "\n" |
26 | R"==(#define CONCAT2(a, b) CONCAt2(a, b) )==" "\n" |
27 | R"==(#define CONCAT3(a, b, c) CONCAT2(CONCAT2(a, b), c) )==" "\n" |
28 | R"==(#if defined(DT_F16) || defined(SRC_DT_F16) || defined(SRC0_DT_F16) \ )==" "\n" |
29 | R"==(|| defined(SRC1_DT_F16) || defined(DST_DT_F16) || defined(WEI_DT_F16) \ )==" "\n" |
30 | R"==(|| defined(BIA_DT_F16) || defined(ACC_DT_F16) )==" "\n" |
31 | R"==(#pragma OPENCL EXTENSION cl_khr_fp16 : enable )==" "\n" |
32 | R"==(#endif )==" "\n" |
33 | R"==(#if DT_F64 )==" "\n" |
34 | R"==(#pragma OPENCL EXTENSION cl_khr_fp64 : enable )==" "\n" |
35 | R"==(#endif )==" "\n" |
36 | R"==(#if DT_F32 == 1 )==" "\n" |
37 | R"==(#define DATA_T float )==" "\n" |
38 | R"==(#define DATA2_T float2 )==" "\n" |
39 | R"==(#define DATA4_T float4 )==" "\n" |
40 | R"==(#define DATA8_T float8 )==" "\n" |
41 | R"==(#define DATA16_T float16 )==" "\n" |
42 | R"==(#define DATA_MAX FLT_MAX )==" "\n" |
43 | R"==(#define DATA_MIN -DATA_MAX )==" "\n" |
44 | R"==(#define DATA_ZERO 0.0f )==" "\n" |
45 | R"==(#define DATA_ONE 1.0f )==" "\n" |
46 | R"==(#define DEF_ACC_DATA_T float )==" "\n" |
47 | R"==(#define DEF_ACC_DATA2_T float2 )==" "\n" |
48 | R"==(#define DEF_ACC_DATA4_T float4 )==" "\n" |
49 | R"==(#define DEF_ACC_DATA8_T float8 )==" "\n" |
50 | R"==(#define POST_OP_DATA_T float )==" "\n" |
51 | R"==(#define TO_DATA_T(v) (float)(v) )==" "\n" |
52 | R"==(#define TO_DEF_ACC_DATA_T convert_float )==" "\n" |
53 | R"==(#define TO_DEF_ACC_DATA2_T convert_float2 )==" "\n" |
54 | R"==(#define TO_DEF_ACC_DATA4_T convert_float4 )==" "\n" |
55 | R"==(#define TO_DEF_ACC_DATA8_T convert_float8 )==" "\n" |
56 | R"==(#define DATA_TO_REF convert_float )==" "\n" |
57 | R"==(#define CONVERT_DATA_T convert_float )==" "\n" |
58 | R"==(#define CONVERT_DATA2_T convert_float2 )==" "\n" |
59 | R"==(#define CONVERT_DATA4_T convert_float4 )==" "\n" |
60 | R"==(#define CONVERT_DATA8_T convert_float8 )==" "\n" |
61 | R"==(#define CONVERT_FLOAT_T convert_float )==" "\n" |
62 | R"==(#define CONVERT_FLOAT2_T convert_float2 )==" "\n" |
63 | R"==(#define CONVERT_FLOAT4_T convert_float4 )==" "\n" |
64 | R"==(#define CONVERT_FLOAT8_T convert_float8 )==" "\n" |
65 | R"==(#define BLOCK_READ intel_sub_group_block_read )==" "\n" |
66 | R"==(#define BLOCK_WRITE intel_sub_group_block_write )==" "\n" |
67 | R"==(#define BLOCK_READ2 intel_sub_group_block_read2 )==" "\n" |
68 | R"==(#define BLOCK_READ4 intel_sub_group_block_read4 )==" "\n" |
69 | R"==(#define BLOCK_READ8 intel_sub_group_block_read8 )==" "\n" |
70 | R"==(#define BLOCK_WRITE2 intel_sub_group_block_write2 )==" "\n" |
71 | R"==(#define BLOCK_WRITE4 intel_sub_group_block_write4 )==" "\n" |
72 | R"==(#define BLOCK_WRITE8 intel_sub_group_block_write8 )==" "\n" |
73 | R"==(#define AS_DATA_T as_float )==" "\n" |
74 | R"==(#define AS_DATA2_T as_float2 )==" "\n" |
75 | R"==(#define AS_DATA4_T as_float4 )==" "\n" |
76 | R"==(#define AS_DATA8_T as_float8 )==" "\n" |
77 | R"==(#define AS_UINT_T as_uint )==" "\n" |
78 | R"==(#define AS_UINT2_T as_uint2 )==" "\n" |
79 | R"==(#define AS_UINT4_T as_uint4 )==" "\n" |
80 | R"==(#define AS_UINT8_T as_uint8 )==" "\n" |
81 | R"==(#define BLOCK_DATA_T uint )==" "\n" |
82 | R"==(#define BLOCK_DATA2_T uint2 )==" "\n" |
83 | R"==(#define BLOCK_DATA4_T uint4 )==" "\n" |
84 | R"==(#define BLOCK_DATA8_T uint8 )==" "\n" |
85 | R"==(#define AS_BLOCK_DATA_T as_uint )==" "\n" |
86 | R"==(#define AS_BLOCK_DATA2_T as_uint2 )==" "\n" |
87 | R"==(#define AS_BLOCK_DATA4_T as_uint4 )==" "\n" |
88 | R"==(#define AS_BLOCK_DATA8_T as_uint8 )==" "\n" |
89 | R"==(#define MMAD_DATA_T uint )==" "\n" |
90 | R"==(#define MMAD_DATA4_T uint4 )==" "\n" |
91 | R"==(#define MMAD_DATA8_T uint8 )==" "\n" |
92 | R"==(#define MMAD_ACC_DATA4_T float4 )==" "\n" |
93 | R"==(#define MMAD_ACC_DATA8_T float8 )==" "\n" |
94 | R"==(#elif DT_F64 == 1 )==" "\n" |
95 | R"==(#define DATA_T double )==" "\n" |
96 | R"==(#define DATA2_T double2 )==" "\n" |
97 | R"==(#define DATA4_T double4 )==" "\n" |
98 | R"==(#define DATA8_T double8 )==" "\n" |
99 | R"==(#define DATA16_T double16 )==" "\n" |
100 | R"==(#define DATA_MAX DBL_MAX )==" "\n" |
101 | R"==(#define DATA_MIN -DATA_MAX )==" "\n" |
102 | R"==(#define DATA_ZERO 0.0d )==" "\n" |
103 | R"==(#define DATA_ONE 1.0d )==" "\n" |
104 | R"==(#define DEF_ACC_DATA_T double )==" "\n" |
105 | R"==(#define DEF_ACC_DATA2_T double2 )==" "\n" |
106 | R"==(#define DEF_ACC_DATA4_T double4 )==" "\n" |
107 | R"==(#define DEF_ACC_DATA8_T double8 )==" "\n" |
108 | R"==(#define POST_OP_DATA_T double )==" "\n" |
109 | R"==(#define TO_DATA_T(v) (double)(v) )==" "\n" |
110 | R"==(#define TO_DEF_ACC_DATA_T(v) (double)(v) )==" "\n" |
111 | R"==(#define DATA_TO_REF convert_float )==" "\n" |
112 | R"==(#define CONVERT_DATA_T convert_double )==" "\n" |
113 | R"==(#define CONVERT_DATA2_T convert_double2 )==" "\n" |
114 | R"==(#define CONVERT_DATA4_T convert_double4 )==" "\n" |
115 | R"==(#define CONVERT_DATA8_T convert_double8 )==" "\n" |
116 | R"==(#define CONVERT_FLOAT_T convert_float )==" "\n" |
117 | R"==(#define CONVERT_FLOAT2_T convert_float2 )==" "\n" |
118 | R"==(#define CONVERT_FLOAT4_T convert_float4 )==" "\n" |
119 | R"==(#define CONVERT_FLOAT8_T convert_float8 )==" "\n" |
120 | R"==(#define AS_DATA_T as_double )==" "\n" |
121 | R"==(#define AS_DATA2_T as_double2 )==" "\n" |
122 | R"==(#define AS_DATA4_T as_double4 )==" "\n" |
123 | R"==(#define AS_DATA8_T as_double8 )==" "\n" |
124 | R"==(#elif DT_F16 == 1 )==" "\n" |
125 | R"==(#define DATA_T half )==" "\n" |
126 | R"==(#define DATA2_T half2 )==" "\n" |
127 | R"==(#define DATA4_T half4 )==" "\n" |
128 | R"==(#define DATA8_T half8 )==" "\n" |
129 | R"==(#define DATA16_T half16 )==" "\n" |
130 | R"==(#define AS_DATA2_T as_half2 )==" "\n" |
131 | R"==(#define DATA_MAX HALF_MAX )==" "\n" |
132 | R"==(#define DATA_MIN -DATA_MAX )==" "\n" |
133 | R"==(#define DATA_ZERO 0.0h )==" "\n" |
134 | R"==(#define DATA_ONE 1.0h )==" "\n" |
135 | R"==(#define DEF_ACC_DATA_T float )==" "\n" |
136 | R"==(#define DEF_ACC_DATA2_T float2 )==" "\n" |
137 | R"==(#define DEF_ACC_DATA4_T float4 )==" "\n" |
138 | R"==(#define DEF_ACC_DATA8_T float8 )==" "\n" |
139 | R"==(#define POST_OP_DATA_T float )==" "\n" |
140 | R"==(#define TO_DATA_T convert_half )==" "\n" |
141 | R"==(#define TO_DEF_ACC_DATA_T convert_float )==" "\n" |
142 | R"==(#define TO_DEF_ACC_DATA2_T convert_float2 )==" "\n" |
143 | R"==(#define TO_DEF_ACC_DATA4_T convert_float4 )==" "\n" |
144 | R"==(#define TO_DEF_ACC_DATA8_T convert_float8 )==" "\n" |
145 | R"==(#define DATA_TO_REF convert_half )==" "\n" |
146 | R"==(#define CONVERT_DATA_T convert_half )==" "\n" |
147 | R"==(#define CONVERT_DATA2_T convert_half2 )==" "\n" |
148 | R"==(#define CONVERT_DATA4_T convert_half4 )==" "\n" |
149 | R"==(#define CONVERT_DATA8_T convert_half8 )==" "\n" |
150 | R"==(#define CONVERT_FLOAT_T convert_float )==" "\n" |
151 | R"==(#define CONVERT_FLOAT2_T convert_float2 )==" "\n" |
152 | R"==(#define CONVERT_FLOAT4_T convert_float4 )==" "\n" |
153 | R"==(#define CONVERT_FLOAT8_T convert_float8 )==" "\n" |
154 | R"==(#define BLOCK_READ intel_sub_group_block_read_us )==" "\n" |
155 | R"==(#define BLOCK_WRITE intel_sub_group_block_write_us )==" "\n" |
156 | R"==(#define BLOCK_READ2 intel_sub_group_block_read_us2 )==" "\n" |
157 | R"==(#define BLOCK_READ4 intel_sub_group_block_read_us4 )==" "\n" |
158 | R"==(#define BLOCK_READ8 intel_sub_group_block_read_us8 )==" "\n" |
159 | R"==(#define BLOCK_WRITE2 intel_sub_group_block_write_us2 )==" "\n" |
160 | R"==(#define BLOCK_WRITE4 intel_sub_group_block_write_us4 )==" "\n" |
161 | R"==(#define BLOCK_WRITE8 intel_sub_group_block_write_us8 )==" "\n" |
162 | R"==(#define AS_DATA_T as_half )==" "\n" |
163 | R"==(#define AS_DATA2_T as_half2 )==" "\n" |
164 | R"==(#define AS_DATA4_T as_half4 )==" "\n" |
165 | R"==(#define AS_DATA8_T as_half8 )==" "\n" |
166 | R"==(#define AS_UINT_T as_ushort )==" "\n" |
167 | R"==(#define AS_UINT2_T as_ushort2 )==" "\n" |
168 | R"==(#define AS_UINT4_T as_ushort4 )==" "\n" |
169 | R"==(#define AS_UINT8_T as_ushort8 )==" "\n" |
170 | R"==(#define BLOCK_DATA_T ushort )==" "\n" |
171 | R"==(#define BLOCK_DATA2_T ushort2 )==" "\n" |
172 | R"==(#define BLOCK_DATA4_T ushort4 )==" "\n" |
173 | R"==(#define BLOCK_DATA8_T ushort8 )==" "\n" |
174 | R"==(#define AS_BLOCK_DATA_T as_ushort )==" "\n" |
175 | R"==(#define AS_BLOCK_DATA2_T as_ushort2 )==" "\n" |
176 | R"==(#define AS_BLOCK_DATA4_T as_ushort4 )==" "\n" |
177 | R"==(#define AS_BLOCK_DATA8_T as_ushort8 )==" "\n" |
178 | R"==(#define MMAD_DATA_T uint )==" "\n" |
179 | R"==(#define MMAD_DATA4_T uint4 )==" "\n" |
180 | R"==(#define MMAD_DATA8_T uint8 )==" "\n" |
181 | R"==(#define MMAD_ACC_DATA4_T float4 )==" "\n" |
182 | R"==(#define MMAD_ACC_DATA8_T float8 )==" "\n" |
183 | R"==(#elif DT_BF16 == 1 )==" "\n" |
184 | R"==(#define DATA_T ushort )==" "\n" |
185 | R"==(#define DATA2_T ushort2 )==" "\n" |
186 | R"==(#define POST_OP_DATA_T float )==" "\n" |
187 | R"==(#define DATA2_T ushort2 )==" "\n" |
188 | R"==(#define DATA4_T ushort4 )==" "\n" |
189 | R"==(#define DATA8_T ushort8 )==" "\n" |
190 | R"==(#define DATA16_T ushort16 )==" "\n" |
191 | R"==(#define DATA_MAX (ushort)0x7F7F )==" "\n" |
192 | R"==(#define DATA_MIN (ushort)0xFF7F )==" "\n" |
193 | R"==(#define DATA_ZERO (ushort)0x0000 )==" "\n" |
194 | R"==(#define DATA_ONE (ushort)0x3F80 )==" "\n" |
195 | R"==(#define DEF_ACC_DATA_T float )==" "\n" |
196 | R"==(#define DEF_ACC_DATA2_T float2 )==" "\n" |
197 | R"==(#define DEF_ACC_DATA4_T float4 )==" "\n" |
198 | R"==(#define DEF_ACC_DATA8_T float8 )==" "\n" |
199 | R"==(#define TO_DATA_T cvt_f32_to_bf16 )==" "\n" |
200 | R"==(#define TO_DEF_ACC_DATA_T cvt_bf16_to_f32 )==" "\n" |
201 | R"==(#define TO_DEF_ACC_DATA2_T cvt_bf16_to_f32 )==" "\n" |
202 | R"==(#define TO_DEF_ACC_DATA4_T cvt_bf16_to_f32 )==" "\n" |
203 | R"==(#define TO_DEF_ACC_DATA8_T cvt_bf16_to_f32 )==" "\n" |
204 | R"==(#define DATA_TO_REF cvt_bf16_to_f32 )==" "\n" |
205 | R"==(#define CONVERT_DATA_T(v) cvt_f32_to_bf16(convert_float(v)) )==" "\n" |
206 | R"==(#define CONVERT_DATA2_T(v) cvt_f32_to_bf16(convert_float2(v)) )==" "\n" |
207 | R"==(#define CONVERT_DATA4_T(v) cvt_f32_to_bf16(convert_float4(v)) )==" "\n" |
208 | R"==(#define CONVERT_DATA8_T(v) cvt_f32_to_bf16(convert_float8(v)) )==" "\n" |
209 | R"==(#define CONVERT_FLOAT_T cvt_bf16_to_f32 )==" "\n" |
210 | R"==(#define CONVERT_FLOAT2_T cvt_bf16_to_f32 )==" "\n" |
211 | R"==(#define CONVERT_FLOAT4_T cvt_bf16_to_f32 )==" "\n" |
212 | R"==(#define CONVERT_FLOAT8_T cvt_bf16_to_f32 )==" "\n" |
213 | R"==(#define BLOCK_READ intel_sub_group_block_read_us )==" "\n" |
214 | R"==(#define BLOCK_WRITE intel_sub_group_block_write_us )==" "\n" |
215 | R"==(#define BLOCK_READ2 intel_sub_group_block_read_us2 )==" "\n" |
216 | R"==(#define BLOCK_READ4 intel_sub_group_block_read_us4 )==" "\n" |
217 | R"==(#define BLOCK_READ8 intel_sub_group_block_read_us8 )==" "\n" |
218 | R"==(#define BLOCK_WRITE2 intel_sub_group_block_write_us2 )==" "\n" |
219 | R"==(#define BLOCK_WRITE4 intel_sub_group_block_write_us4 )==" "\n" |
220 | R"==(#define BLOCK_WRITE8 intel_sub_group_block_write_us8 )==" "\n" |
221 | R"==(#define AS_DATA_T as_ushort )==" "\n" |
222 | R"==(#define AS_DATA2_T as_ushort2 )==" "\n" |
223 | R"==(#define AS_DATA4_T as_ushort4 )==" "\n" |
224 | R"==(#define AS_DATA8_T as_ushort8 )==" "\n" |
225 | R"==(#define AS_UINT_T as_ushort )==" "\n" |
226 | R"==(#define AS_UINT2_T as_ushort2 )==" "\n" |
227 | R"==(#define AS_UINT4_T as_ushort4 )==" "\n" |
228 | R"==(#define AS_UINT8_T as_ushort8 )==" "\n" |
229 | R"==(#define BLOCK_DATA_T ushort )==" "\n" |
230 | R"==(#define BLOCK_DATA2_T ushort2 )==" "\n" |
231 | R"==(#define BLOCK_DATA4_T ushort4 )==" "\n" |
232 | R"==(#define BLOCK_DATA8_T ushort8 )==" "\n" |
233 | R"==(#define AS_BLOCK_DATA_T as_ushort )==" "\n" |
234 | R"==(#define AS_BLOCK_DATA2_T as_ushort2 )==" "\n" |
235 | R"==(#define AS_BLOCK_DATA4_T as_ushort4 )==" "\n" |
236 | R"==(#define AS_BLOCK_DATA8_T as_ushort8 )==" "\n" |
237 | R"==(#define MMAD_DATA_T uint )==" "\n" |
238 | R"==(#define MMAD_DATA4_T uint4 )==" "\n" |
239 | R"==(#define MMAD_DATA8_T uint8 )==" "\n" |
240 | R"==(#define MMAD_ACC_DATA4_T float4 )==" "\n" |
241 | R"==(#define MMAD_ACC_DATA8_T float8 )==" "\n" |
242 | R"==(#elif DT_S8 == 1 )==" "\n" |
243 | R"==(#define DATA_T char )==" "\n" |
244 | R"==(#define DATA2_T char2 )==" "\n" |
245 | R"==(#define DATA4_T char4 )==" "\n" |
246 | R"==(#define DATA8_T char8 )==" "\n" |
247 | R"==(#define DATA16_T char16 )==" "\n" |
248 | R"==(#define DATA_MAX CHAR_MAX )==" "\n" |
249 | R"==(#define DATA_MIN CHAR_MIN )==" "\n" |
250 | R"==(#define DATA_ZERO 0 )==" "\n" |
251 | R"==(#define DATA_ONE 1 )==" "\n" |
252 | R"==(#define INT8_T int8 )==" "\n" |
253 | R"==(#define DEF_ACC_DATA_T int )==" "\n" |
254 | R"==(#define DEF_ACC_DATA2_T int2 )==" "\n" |
255 | R"==(#define DEF_ACC_DATA4_T int4 )==" "\n" |
256 | R"==(#define DEF_ACC_DATA8_T int8 )==" "\n" |
257 | R"==(#define POST_OP_DATA_T float )==" "\n" |
258 | R"==(#define TO_DATA_T(v) convert_char_sat_rte(v) )==" "\n" |
259 | R"==(#define TO_DEF_ACC_DATA_T convert_int_sat_rte )==" "\n" |
260 | R"==(#define TO_DEF_ACC_DATA2_T convert_int2_sat_rte )==" "\n" |
261 | R"==(#define TO_DEF_ACC_DATA4_T convert_int4_sat_rte )==" "\n" |
262 | R"==(#define TO_DEF_ACC_DATA8_T convert_int8_sat_rte )==" "\n" |
263 | R"==(#define DATA_TO_REF convert_float )==" "\n" |
264 | R"==(#define CONVERT_DATA_T convert_char_sat_rte )==" "\n" |
265 | R"==(#define CONVERT_DATA2_T convert_char2_sat_rte )==" "\n" |
266 | R"==(#define CONVERT_DATA4_T convert_char4_sat_rte )==" "\n" |
267 | R"==(#define CONVERT_DATA8_T convert_char8_sat_rte )==" "\n" |
268 | R"==(#define CONVERT_FLOAT_T convert_float )==" "\n" |
269 | R"==(#define CONVERT_FLOAT2_T convert_float2 )==" "\n" |
270 | R"==(#define CONVERT_FLOAT4_T convert_float4 )==" "\n" |
271 | R"==(#define CONVERT_FLOAT8_T convert_float8 )==" "\n" |
272 | R"==(#define BLOCK_READ intel_sub_group_block_read_uc )==" "\n" |
273 | R"==(#define BLOCK_WRITE intel_sub_group_block_write_uc )==" "\n" |
274 | R"==(#define BLOCK_READ2 intel_sub_group_block_read_uc2 )==" "\n" |
275 | R"==(#define BLOCK_READ4 intel_sub_group_block_read_uc4 )==" "\n" |
276 | R"==(#define BLOCK_READ8 intel_sub_group_block_read_uc8 )==" "\n" |
277 | R"==(#define BLOCK_WRITE2 intel_sub_group_block_write_uc2 )==" "\n" |
278 | R"==(#define BLOCK_WRITE4 intel_sub_group_block_write_uc4 )==" "\n" |
279 | R"==(#define BLOCK_WRITE8 intel_sub_group_block_write_uc8 )==" "\n" |
280 | R"==(#define AS_DATA_T as_char )==" "\n" |
281 | R"==(#define AS_DATA2_T as_char2 )==" "\n" |
282 | R"==(#define AS_DATA4_T as_char4 )==" "\n" |
283 | R"==(#define AS_DATA8_T as_char8 )==" "\n" |
284 | R"==(#define AS_DATA16_T as_char16 )==" "\n" |
285 | R"==(#define AS_UINT_T as_uchar )==" "\n" |
286 | R"==(#define AS_UINT2_T as_uchar2 )==" "\n" |
287 | R"==(#define AS_UINT4_T as_uchar4 )==" "\n" |
288 | R"==(#define AS_UINT8_T as_uchar8 )==" "\n" |
289 | R"==(#define AS_INT8_T as_int8 )==" "\n" |
290 | R"==(#define BLOCK_DATA_T uchar )==" "\n" |
291 | R"==(#define BLOCK_DATA2_T uchar2 )==" "\n" |
292 | R"==(#define BLOCK_DATA4_T uchar4 )==" "\n" |
293 | R"==(#define BLOCK_DATA8_T uchar8 )==" "\n" |
294 | R"==(#define AS_BLOCK_DATA_T as_uchar )==" "\n" |
295 | R"==(#define AS_BLOCK_DATA2_T as_uchar2 )==" "\n" |
296 | R"==(#define AS_BLOCK_DATA4_T as_uchar4 )==" "\n" |
297 | R"==(#define AS_BLOCK_DATA8_T as_uchar8 )==" "\n" |
298 | R"==(#define MMAD_DATA_T int )==" "\n" |
299 | R"==(#define MMAD_DATA4_T int4 )==" "\n" |
300 | R"==(#define MMAD_DATA8_T int8 )==" "\n" |
301 | R"==(#define MMAD_ACC_DATA4_T int4 )==" "\n" |
302 | R"==(#define MMAD_ACC_DATA8_T int8 )==" "\n" |
303 | R"==(#elif DT_U8 == 1 )==" "\n" |
304 | R"==(#define DATA_T uchar )==" "\n" |
305 | R"==(#define DATA2_T uchar2 )==" "\n" |
306 | R"==(#define DATA4_T uchar4 )==" "\n" |
307 | R"==(#define DATA8_T uchar8 )==" "\n" |
308 | R"==(#define DATA16_T uchar16 )==" "\n" |
309 | R"==(#define DATA_MAX UCHAR_MAX )==" "\n" |
310 | R"==(#define DATA_MIN 0 )==" "\n" |
311 | R"==(#define DATA_ZERO 0 )==" "\n" |
312 | R"==(#define DATA_ONE 1 )==" "\n" |
313 | R"==(#define INT8_T uint8 )==" "\n" |
314 | R"==(#define DEF_ACC_DATA_T int )==" "\n" |
315 | R"==(#define DEF_ACC_DATA2_T int2 )==" "\n" |
316 | R"==(#define DEF_ACC_DATA4_T int4 )==" "\n" |
317 | R"==(#define DEF_ACC_DATA8_T int8 )==" "\n" |
318 | R"==(#define POST_OP_DATA_T float )==" "\n" |
319 | R"==(#define TO_DATA_T(v) convert_uchar_sat_rte(v) )==" "\n" |
320 | R"==(#define TO_DEF_ACC_DATA_T convert_int_sat_rte )==" "\n" |
321 | R"==(#define TO_DEF_ACC_DATA2_T convert_int2_sat_rte )==" "\n" |
322 | R"==(#define TO_DEF_ACC_DATA4_T convert_int4_sat_rte )==" "\n" |
323 | R"==(#define TO_DEF_ACC_DATA8_T convert_int8_sat_rte )==" "\n" |
324 | R"==(#define DATA_TO_REF convert_float )==" "\n" |
325 | R"==(#define CONVERT_DATA_T convert_uchar_sat_rte )==" "\n" |
326 | R"==(#define CONVERT_DATA2_T convert_uchar2_sat_rte )==" "\n" |
327 | R"==(#define CONVERT_DATA4_T convert_uchar4_sat_rte )==" "\n" |
328 | R"==(#define CONVERT_DATA8_T convert_uchar8_sat_rte )==" "\n" |
329 | R"==(#define CONVERT_FLOAT_T convert_float )==" "\n" |
330 | R"==(#define CONVERT_FLOAT2_T convert_float2 )==" "\n" |
331 | R"==(#define CONVERT_FLOAT4_T convert_float4 )==" "\n" |
332 | R"==(#define CONVERT_FLOAT8_T convert_float8 )==" "\n" |
333 | R"==(#define BLOCK_READ intel_sub_group_block_read_uc )==" "\n" |
334 | R"==(#define BLOCK_WRITE intel_sub_group_block_write_uc )==" "\n" |
335 | R"==(#define BLOCK_READ2 intel_sub_group_block_read_uc2 )==" "\n" |
336 | R"==(#define BLOCK_READ4 intel_sub_group_block_read_uc4 )==" "\n" |
337 | R"==(#define BLOCK_READ8 intel_sub_group_block_read_uc8 )==" "\n" |
338 | R"==(#define BLOCK_WRITE2 intel_sub_group_block_write_uc2 )==" "\n" |
339 | R"==(#define BLOCK_WRITE4 intel_sub_group_block_write_uc4 )==" "\n" |
340 | R"==(#define BLOCK_WRITE8 intel_sub_group_block_write_uc8 )==" "\n" |
341 | R"==(#define AS_DATA_T as_uchar )==" "\n" |
342 | R"==(#define AS_DATA2_T as_uchar2 )==" "\n" |
343 | R"==(#define AS_DATA4_T as_uchar4 )==" "\n" |
344 | R"==(#define AS_DATA8_T as_uchar8 )==" "\n" |
345 | R"==(#define AS_DATA16_T as_uchar16 )==" "\n" |
346 | R"==(#define AS_UINT_T as_uchar )==" "\n" |
347 | R"==(#define AS_UINT2_T as_uchar2 )==" "\n" |
348 | R"==(#define AS_UINT4_T as_uchar4 )==" "\n" |
349 | R"==(#define AS_UINT8_T as_uchar8 )==" "\n" |
350 | R"==(#define AS_INT8_T as_uint8 )==" "\n" |
351 | R"==(#define BLOCK_DATA_T uchar )==" "\n" |
352 | R"==(#define BLOCK_DATA2_T uchar2 )==" "\n" |
353 | R"==(#define BLOCK_DATA4_T uchar4 )==" "\n" |
354 | R"==(#define BLOCK_DATA8_T uchar8 )==" "\n" |
355 | R"==(#define AS_BLOCK_DATA_T as_uchar )==" "\n" |
356 | R"==(#define AS_BLOCK_DATA2_T as_uchar2 )==" "\n" |
357 | R"==(#define AS_BLOCK_DATA4_T as_uchar4 )==" "\n" |
358 | R"==(#define AS_BLOCK_DATA8_T as_uchar8 )==" "\n" |
359 | R"==(#define MMAD_DATA_T uint )==" "\n" |
360 | R"==(#define MMAD_DATA4_T uint4 )==" "\n" |
361 | R"==(#define MMAD_DATA8_T uint8 )==" "\n" |
362 | R"==(#define MMAD_ACC_DATA4_T int4 )==" "\n" |
363 | R"==(#define MMAD_ACC_DATA8_T int8 )==" "\n" |
364 | R"==(#elif DT_S32 == 1 )==" "\n" |
365 | R"==(#define MMAD_DATA_T uint )==" "\n" |
366 | R"==(#define MMAD_DATA4_T uint4 )==" "\n" |
367 | R"==(#define MMAD_DATA8_T uint8 )==" "\n" |
368 | R"==(#define DATA_T int )==" "\n" |
369 | R"==(#define DATA2_T int2 )==" "\n" |
370 | R"==(#define DATA4_T int4 )==" "\n" |
371 | R"==(#define DATA8_T int8 )==" "\n" |
372 | R"==(#define DATA16_T int16 )==" "\n" |
373 | R"==(#define DATA_MAX INT_MAX )==" "\n" |
374 | R"==(#define DATA_MIN INT_MIN )==" "\n" |
375 | R"==(#define DATA_ZERO 0 )==" "\n" |
376 | R"==(#define DATA_ONE 1 )==" "\n" |
377 | R"==(#define DATA_TO_REF convert_float )==" "\n" |
378 | R"==(#define TO_DATA_T(v) convert_int_sat_rte )==" "\n" |
379 | R"==(#define TO_DATA8_T(v) convert_int8_sat_rte )==" "\n" |
380 | R"==(#define CONVERT_DATA_T convert_int_sat_rte )==" "\n" |
381 | R"==(#define CONVERT_DATA2_T convert_int2_sat_rte )==" "\n" |
382 | R"==(#define CONVERT_DATA4_T convert_int4_sat_rte )==" "\n" |
383 | R"==(#define CONVERT_DATA8_T convert_int8_sat_rte )==" "\n" |
384 | R"==(#define CONVERT_FLOAT_T convert_float )==" "\n" |
385 | R"==(#define CONVERT_FLOAT2_T convert_float2 )==" "\n" |
386 | R"==(#define CONVERT_FLOAT4_T convert_float4 )==" "\n" |
387 | R"==(#define CONVERT_FLOAT8_T convert_float8 )==" "\n" |
388 | R"==(#define DEF_ACC_DATA_T int )==" "\n" |
389 | R"==(#define DEF_ACC_DATA2_T int2 )==" "\n" |
390 | R"==(#define DEF_ACC_DATA4_T int4 )==" "\n" |
391 | R"==(#define DEF_ACC_DATA8_T int8 )==" "\n" |
392 | R"==(#define TO_DATA_T(v) convert_int_sat_rte(v) )==" "\n" |
393 | R"==(#define TO_DEF_ACC_DATA_T convert_int_sat_rte )==" "\n" |
394 | R"==(#define TO_DEF_ACC_DATA2_T convert_int2_sat_rte )==" "\n" |
395 | R"==(#define TO_DEF_ACC_DATA4_T convert_int4_sat_rte )==" "\n" |
396 | R"==(#define TO_DEF_ACC_DATA8_T convert_int8_sat_rte )==" "\n" |
397 | R"==(#define POST_OP_DATA_T float )==" "\n" |
398 | R"==(#define DATA_MIN INT_MIN )==" "\n" |
399 | R"==(#define DATA_MAX INT_MAX )==" "\n" |
400 | R"==(#define DATA_ZERO 0 )==" "\n" |
401 | R"==(#define ROUND )==" "\n" |
402 | R"==(#define BLOCK_READ intel_sub_group_block_read )==" "\n" |
403 | R"==(#define BLOCK_WRITE intel_sub_group_block_write )==" "\n" |
404 | R"==(#define BLOCK_READ2 intel_sub_group_block_read2 )==" "\n" |
405 | R"==(#define BLOCK_READ4 intel_sub_group_block_read4 )==" "\n" |
406 | R"==(#define BLOCK_READ8 intel_sub_group_block_read8 )==" "\n" |
407 | R"==(#define BLOCK_WRITE2 intel_sub_group_block_write2 )==" "\n" |
408 | R"==(#define BLOCK_WRITE4 intel_sub_group_block_write4 )==" "\n" |
409 | R"==(#define BLOCK_WRITE8 intel_sub_group_block_write8 )==" "\n" |
410 | R"==(#define AS_DATA_T as_int )==" "\n" |
411 | R"==(#define AS_DATA2_T as_int2 )==" "\n" |
412 | R"==(#define AS_DATA4_T as_int4 )==" "\n" |
413 | R"==(#define AS_DATA8_T as_int8 )==" "\n" |
414 | R"==(#define AS_UINT_T as_uint )==" "\n" |
415 | R"==(#define AS_UINT2_T as_uint2 )==" "\n" |
416 | R"==(#define AS_UINT4_T as_uint4 )==" "\n" |
417 | R"==(#define AS_UINT8_T as_uint8 )==" "\n" |
418 | R"==(#define BLOCK_DATA_T uint )==" "\n" |
419 | R"==(#define BLOCK_DATA2_T uint2 )==" "\n" |
420 | R"==(#define BLOCK_DATA4_T uint4 )==" "\n" |
421 | R"==(#define BLOCK_DATA8_T uint8 )==" "\n" |
422 | R"==(#define AS_BLOCK_DATA_T as_uint )==" "\n" |
423 | R"==(#define AS_BLOCK_DATA2_T as_uint2 )==" "\n" |
424 | R"==(#define AS_BLOCK_DATA4_T as_uint4 )==" "\n" |
425 | R"==(#define AS_BLOCK_DATA8_T as_uint8 )==" "\n" |
426 | R"==(#elif !defined(DT_UNDEF) )==" "\n" |
427 | R"==(#error "Unexpected data type" )==" "\n" |
428 | R"==(#endif )==" "\n" |
429 | R"==(#if VECT_DT_N == 1 )==" "\n" |
430 | R"==(#define VECT_DATA_T DATA_T )==" "\n" |
431 | R"==(#define VECT_DEF_ACC_DATA_T DEF_ACC_DATA_T )==" "\n" |
432 | R"==(#define AS_VECT_DEF_ACC_DATA_T TO_DEF_ACC_DATA_T )==" "\n" |
433 | R"==(#define AS_VECT_DATA_T AS_DATA_T )==" "\n" |
434 | R"==(#define VECT_BLOCK_READ BLOCK_READ )==" "\n" |
435 | R"==(#define VECT_BLOCK_WRITE BLOCK_WRITE )==" "\n" |
436 | R"==(#define VECT_UINT_READ intel_sub_group_block_read )==" "\n" |
437 | R"==(#define VECT_UINT_WRITE intel_sub_group_block_write )==" "\n" |
438 | R"==(#define VECT_UCHAR_READ intel_sub_group_block_read_uc )==" "\n" |
439 | R"==(#define VECT_UCHAR_WRITE intel_sub_group_block_write_uc )==" "\n" |
440 | R"==(#define VECT_BLOCK_DATA_T BLOCK_DATA_T )==" "\n" |
441 | R"==(#define AS_VECT_BLOCK_DATA_T AS_BLOCK_DATA_T )==" "\n" |
442 | R"==(#define CONVERT_VECT_FLOAT_T CONVERT_FLOAT_T )==" "\n" |
443 | R"==(#define CONVERT_VECTOR_DATA_T CONVERT_DATA_T )==" "\n" |
444 | R"==(#define CONVERT_VECT_CHAR_T convert_char )==" "\n" |
445 | R"==(#define CONVERT_VECT_INT_T convert_int )==" "\n" |
446 | R"==(#define VECT_INT_T int )==" "\n" |
447 | R"==(#define VECT_UINT_T uint )==" "\n" |
448 | R"==(#define VECT_FLOAT_T float )==" "\n" |
449 | R"==(#define VECT_CHAR_T char )==" "\n" |
450 | R"==(#define AS_VECT_INT_T as_int )==" "\n" |
451 | R"==(#define AS_VECT_UINT_T as_uint )==" "\n" |
452 | R"==(#define AS_VECT_FLOAT_T as_float )==" "\n" |
453 | R"==(#define AS_VECT_CHAR_T as_char )==" "\n" |
454 | R"==(#define AS_VECT_UCHAR_T as_uchar )==" "\n" |
455 | R"==(#elif VECT_DT_N == 2 )==" "\n" |
456 | R"==(#define VECT_DATA_T DATA2_T )==" "\n" |
457 | R"==(#define VECT_DEF_ACC_DATA_T DEF_ACC_DATA2_T )==" "\n" |
458 | R"==(#define AS_VECT_DEF_ACC_DATA_T TO_DEF_ACC_DATA2_T )==" "\n" |
459 | R"==(#define AS_VECT_DATA_T AS_DATA2_T )==" "\n" |
460 | R"==(#define VECT_BLOCK_READ BLOCK_READ2 )==" "\n" |
461 | R"==(#define VECT_BLOCK_WRITE BLOCK_WRITE2 )==" "\n" |
462 | R"==(#define VECT_UINT_READ intel_sub_group_block_read2 )==" "\n" |
463 | R"==(#define VECT_UINT_WRITE intel_sub_group_block_write2 )==" "\n" |
464 | R"==(#define VECT_UCHAR_READ intel_sub_group_block_read_uc2 )==" "\n" |
465 | R"==(#define VECT_UCHAR_WRITE intel_sub_group_block_write_uc2 )==" "\n" |
466 | R"==(#define VECT_BLOCK_DATA_T BLOCK_DATA2_T )==" "\n" |
467 | R"==(#define AS_VECT_BLOCK_DATA_T AS_BLOCK_DATA2_T )==" "\n" |
468 | R"==(#define CONVERT_VECT_FLOAT_T CONVERT_FLOAT2_T )==" "\n" |
469 | R"==(#define CONVERT_VECTOR_DATA_T CONVERT_DATA2_T )==" "\n" |
470 | R"==(#define CONVERT_VECT_CHAR_T convert_char2 )==" "\n" |
471 | R"==(#define CONVERT_VECT_INT_T convert_int2 )==" "\n" |
472 | R"==(#define VECT_INT_T int2 )==" "\n" |
473 | R"==(#define VECT_UINT_T uint2 )==" "\n" |
474 | R"==(#define VECT_FLOAT_T float2 )==" "\n" |
475 | R"==(#define VECT_CHAR_T char2 )==" "\n" |
476 | R"==(#define AS_VECT_INT_T as_int2 )==" "\n" |
477 | R"==(#define AS_VECT_UINT_T as_uint2 )==" "\n" |
478 | R"==(#define AS_VECT_FLOAT_T as_float2 )==" "\n" |
479 | R"==(#define AS_VECT_CHAR_T as_char2 )==" "\n" |
480 | R"==(#define AS_VECT_UCHAR_T as_uchar2 )==" "\n" |
481 | R"==(#elif VECT_DT_N == 4 )==" "\n" |
482 | R"==(#define VECT_DATA_T DATA4_T )==" "\n" |
483 | R"==(#define VECT_DEF_ACC_DATA_T DEF_ACC_DATA4_T )==" "\n" |
484 | R"==(#define AS_VECT_DEF_ACC_DATA_T TO_DEF_ACC_DATA4_T )==" "\n" |
485 | R"==(#define AS_VECT_DATA_T AS_DATA4_T )==" "\n" |
486 | R"==(#define VECT_BLOCK_READ BLOCK_READ4 )==" "\n" |
487 | R"==(#define VECT_BLOCK_WRITE BLOCK_WRITE4 )==" "\n" |
488 | R"==(#define VECT_UINT_READ intel_sub_group_block_read4 )==" "\n" |
489 | R"==(#define VECT_UINT_WRITE intel_sub_group_block_write4 )==" "\n" |
490 | R"==(#define VECT_UCHAR_READ intel_sub_group_block_read_uc4 )==" "\n" |
491 | R"==(#define VECT_UCHAR_WRITE intel_sub_group_block_write_uc4 )==" "\n" |
492 | R"==(#define VECT_BLOCK_DATA_T BLOCK_DATA4_T )==" "\n" |
493 | R"==(#define AS_VECT_BLOCK_DATA_T AS_BLOCK_DATA4_T )==" "\n" |
494 | R"==(#define CONVERT_VECT_FLOAT_T CONVERT_FLOAT4_T )==" "\n" |
495 | R"==(#define CONVERT_VECTOR_DATA_T CONVERT_DATA4_T )==" "\n" |
496 | R"==(#define CONVERT_VECT_CHAR_T convert_char4 )==" "\n" |
497 | R"==(#define CONVERT_VECT_INT_T convert_int4 )==" "\n" |
498 | R"==(#define VECT_INT_T int4 )==" "\n" |
499 | R"==(#define VECT_UINT_T uint4 )==" "\n" |
500 | R"==(#define VECT_FLOAT_T float4 )==" "\n" |
501 | R"==(#define VECT_CHAR_T char4 )==" "\n" |
502 | R"==(#define AS_VECT_INT_T as_int4 )==" "\n" |
503 | R"==(#define AS_VECT_UINT_T as_uint4 )==" "\n" |
504 | R"==(#define AS_VECT_FLOAT_T as_float4 )==" "\n" |
505 | R"==(#define AS_VECT_CHAR_T as_char4 )==" "\n" |
506 | R"==(#define AS_VECT_UCHAR_T as_uchar4 )==" "\n" |
507 | R"==(#elif VECT_DT_N == 8 )==" "\n" |
508 | R"==(#define VECT_DATA_T DATA8_T )==" "\n" |
509 | R"==(#define VECT_DEF_ACC_DATA_T DEF_ACC_DATA8_T )==" "\n" |
510 | R"==(#define AS_VECT_DEF_ACC_DATA_T TO_DEF_ACC_DATA8_T )==" "\n" |
511 | R"==(#define AS_VECT_DATA_T AS_DATA8_T )==" "\n" |
512 | R"==(#define VECT_BLOCK_READ BLOCK_READ8 )==" "\n" |
513 | R"==(#define VECT_BLOCK_WRITE BLOCK_WRITE8 )==" "\n" |
514 | R"==(#define VECT_UINT_READ intel_sub_group_block_read8 )==" "\n" |
515 | R"==(#define VECT_UINT_WRITE intel_sub_group_block_write8 )==" "\n" |
516 | R"==(#define VECT_UCHAR_READ intel_sub_group_block_read_uc8 )==" "\n" |
517 | R"==(#define VECT_UCHAR_WRITE intel_sub_group_block_write_uc8 )==" "\n" |
518 | R"==(#define VECT_BLOCK_DATA_T BLOCK_DATA8_T )==" "\n" |
519 | R"==(#define AS_VECT_BLOCK_DATA_T AS_BLOCK_DATA8_T )==" "\n" |
520 | R"==(#define CONVERT_VECT_FLOAT_T CONVERT_FLOAT8_T )==" "\n" |
521 | R"==(#define CONVERT_VECTOR_DATA_T CONVERT_DATA8_T )==" "\n" |
522 | R"==(#define CONVERT_VECT_CHAR_T convert_char8 )==" "\n" |
523 | R"==(#define CONVERT_VECT_INT_T convert_int8 )==" "\n" |
524 | R"==(#define VECT_INT_T int8 )==" "\n" |
525 | R"==(#define VECT_UINT_T uint8 )==" "\n" |
526 | R"==(#define VECT_FLOAT_T float8 )==" "\n" |
527 | R"==(#define VECT_CHAR_T char8 )==" "\n" |
528 | R"==(#define AS_VECT_INT_T as_int8 )==" "\n" |
529 | R"==(#define AS_VECT_UINT_T as_uint8 )==" "\n" |
530 | R"==(#define AS_VECT_FLOAT_T as_float8 )==" "\n" |
531 | R"==(#define AS_VECT_CHAR_T as_char8 )==" "\n" |
532 | R"==(#define AS_VECT_UCHAR_T as_uchar8 )==" "\n" |
533 | R"==(#endif )==" "\n" |
534 | R"==(#define AS_MMAD_DATA_T CONCAT2(as_, MMAD_DATA_T) )==" "\n" |
535 | R"==(#define AS_MMAD_DATA4_T CONCAT2(as_, MMAD_DATA4_T) )==" "\n" |
536 | R"==(#define AS_MMAD_DATA8_T CONCAT2(as_, MMAD_DATA8_T) )==" "\n" |
537 | R"==(#ifdef SRC_DATA_T )==" "\n" |
538 | R"==(#define SRC_DATA2_T CONCAT2(SRC_DATA_T, 2) )==" "\n" |
539 | R"==(#define SRC_DATA4_T CONCAT2(SRC_DATA_T, 4) )==" "\n" |
540 | R"==(#define SRC_DATA8_T CONCAT2(SRC_DATA_T, 8) )==" "\n" |
541 | R"==(#define SRC_DATA16_T CONCAT2(SRC_DATA_T, 16) )==" "\n" |
542 | R"==(#ifdef SRC_DT_U8 )==" "\n" |
543 | R"==(#define SRC_MMAD_DATA_T uint )==" "\n" |
544 | R"==(#define SRC_MMAD_DATA4_T uint4 )==" "\n" |
545 | R"==(#define SRC_MMAD_DATA8_T uint8 )==" "\n" |
546 | R"==(#elif SRC_DT_S8 )==" "\n" |
547 | R"==(#define SRC_MMAD_DATA_T int )==" "\n" |
548 | R"==(#define SRC_MMAD_DATA4_T int4 )==" "\n" |
549 | R"==(#define SRC_MMAD_DATA8_T int8 )==" "\n" |
550 | R"==(#elif SRC_DT_F16 || SRC_DT_BF16 )==" "\n" |
551 | R"==(#define SRC_MMAD_DATA_T uint )==" "\n" |
552 | R"==(#define SRC_MMAD_DATA4_T uint4 )==" "\n" |
553 | R"==(#define SRC_MMAD_DATA8_T uint8 )==" "\n" |
554 | R"==(#endif )==" "\n" |
555 | R"==(#if defined(SRC_DT_U8) || defined(SRC_DT_S8) )==" "\n" |
556 | R"==(#define SRC_MMAD_ACC_DATA4_T int4 )==" "\n" |
557 | R"==(#define SRC_MMAD_ACC_DATA8_T int8 )==" "\n" |
558 | R"==(#else )==" "\n" |
559 | R"==(#define SRC_MMAD_ACC_DATA4_T float4 )==" "\n" |
560 | R"==(#define SRC_MMAD_ACC_DATA8_T float8 )==" "\n" |
561 | R"==(#endif )==" "\n" |
562 | R"==(#define AS_SRC_DATA2_T CONCAT2(as_, SRC_DATA2_T) )==" "\n" |
563 | R"==(#define AS_SRC_DATA4_T CONCAT2(as_, SRC_DATA4_T) )==" "\n" |
564 | R"==(#define AS_SRC_DATA8_T CONCAT2(as_, SRC_DATA8_T) )==" "\n" |
565 | R"==(#define AS_SRC_DATA16_T CONCAT2(as_, SRC_DATA16_T) )==" "\n" |
566 | R"==(#define AS_SRC_MMAD_DATA_T CONCAT2(as_, SRC_MMAD_DATA_T) )==" "\n" |
567 | R"==(#define AS_SRC_MMAD_DATA4_T CONCAT2(as_, SRC_MMAD_DATA4_T) )==" "\n" |
568 | R"==(#define AS_SRC_MMAD_DATA8_T CONCAT2(as_, SRC_MMAD_DATA8_T) )==" "\n" |
569 | R"==(#if SRC_DT_BF16 )==" "\n" |
570 | R"==(#define SRC_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
571 | R"==(#define SRC_TO_REF8(x) cvt_bf16_to_f32(x) )==" "\n" |
572 | R"==(#define REF_TO_SRC(x) cvt_f32_to_bf16(x) )==" "\n" |
573 | R"==(#else )==" "\n" |
574 | R"==(#define SRC_TO_REF(x) (x) )==" "\n" |
575 | R"==(#define SRC_TO_REF8(x) (x) )==" "\n" |
576 | R"==(#define REF_TO_SRC(x) (x) )==" "\n" |
577 | R"==(#endif )==" "\n" |
578 | R"==(#if SRC_DT_BF16 )==" "\n" |
579 | R"==(#define TO_SRC(x) cvt_f32_to_bf16(x) )==" "\n" |
580 | R"==(#elif SRC_DT_U8 )==" "\n" |
581 | R"==(#define TO_SRC(x) convert_uchar_sat_rte(x) )==" "\n" |
582 | R"==(#elif SRC_DT_S8 )==" "\n" |
583 | R"==(#define TO_SRC(x) convert_char_sat_rte(x) )==" "\n" |
584 | R"==(#elif SRC_DT_S32 )==" "\n" |
585 | R"==(#define TO_SRC(x) convert_int_sat_rte(x) )==" "\n" |
586 | R"==(#else )==" "\n" |
587 | R"==(#define TO_SRC(x) (x) )==" "\n" |
588 | R"==(#endif )==" "\n" |
589 | R"==(#endif )==" "\n" |
590 | R"==(#ifdef A_DATA_T )==" "\n" |
591 | R"==(#define A_DATA8_T CONCAT2(A_DATA_T, 8) )==" "\n" |
592 | R"==(#if A_DT_BF16 )==" "\n" |
593 | R"==(#define A_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
594 | R"==(#define A_TO_REF8(x) cvt_bf16_to_f32(x) )==" "\n" |
595 | R"==(#define REF_TO_A(x) cvt_f32_to_bf16(x) )==" "\n" |
596 | R"==(#else )==" "\n" |
597 | R"==(#define A_TO_REF(x) (x) )==" "\n" |
598 | R"==(#define A_TO_REF8(x) (x) )==" "\n" |
599 | R"==(#define REF_TO_A(x) (x) )==" "\n" |
600 | R"==(#endif )==" "\n" |
601 | R"==(#if A_DT_BF16 )==" "\n" |
602 | R"==(#define TO_A(x) cvt_f32_to_bf16(x) )==" "\n" |
603 | R"==(#elif A_DT_U8 )==" "\n" |
604 | R"==(#define TO_A(x) convert_uchar_sat_rte(x) )==" "\n" |
605 | R"==(#elif A_DT_S8 )==" "\n" |
606 | R"==(#define TO_A(x) convert_char_sat_rte(x) )==" "\n" |
607 | R"==(#elif A_DT_S32 )==" "\n" |
608 | R"==(#define TO_A(x) convert_int_sat_rte(x) )==" "\n" |
609 | R"==(#else )==" "\n" |
610 | R"==(#define TO_A(x) (x) )==" "\n" |
611 | R"==(#endif )==" "\n" |
612 | R"==(#endif )==" "\n" |
613 | R"==(#ifdef WEI_DATA_T )==" "\n" |
614 | R"==(#if WEI_DT_BF16 )==" "\n" |
615 | R"==(#define WEI_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
616 | R"==(#define REF_TO_WEI(x) cvt_f32_to_bf16(x) )==" "\n" |
617 | R"==(#else )==" "\n" |
618 | R"==(#define WEI_TO_REF(x) (x) )==" "\n" |
619 | R"==(#define REF_TO_WEI(x) (x) )==" "\n" |
620 | R"==(#endif )==" "\n" |
621 | R"==(#if WEI_DT_BF16 )==" "\n" |
622 | R"==(#define TO_WEI(x) cvt_f32_to_bf16(x) )==" "\n" |
623 | R"==(#elif WEI_DT_U8 )==" "\n" |
624 | R"==(#define TO_WEI(x) convert_uchar_sat_rte(x) )==" "\n" |
625 | R"==(#elif WEI_DT_S8 )==" "\n" |
626 | R"==(#define TO_WEI(x) convert_char_sat_rte(x) )==" "\n" |
627 | R"==(#elif WEI_DT_S32 )==" "\n" |
628 | R"==(#define TO_WEI(x) convert_int_sat_rte(x) )==" "\n" |
629 | R"==(#else )==" "\n" |
630 | R"==(#define TO_WEI(x) (x) )==" "\n" |
631 | R"==(#endif )==" "\n" |
632 | R"==(#endif )==" "\n" |
633 | R"==(#ifdef DIFF_WEI_DATA_T )==" "\n" |
634 | R"==(#if DIFF_WEI_DT_BF16 )==" "\n" |
635 | R"==(#define DIFF_WEI_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
636 | R"==(#define REF_TO_DIFF_WEI(x) cvt_f32_to_bf16(x) )==" "\n" |
637 | R"==(#else )==" "\n" |
638 | R"==(#define DIFF_WEI_TO_REF(x) (x) )==" "\n" |
639 | R"==(#define REF_TO_DIFF_WEI(x) (x) )==" "\n" |
640 | R"==(#endif )==" "\n" |
641 | R"==(#if DIFF_WEI_DT_BF16 )==" "\n" |
642 | R"==(#define TO_DIFF_WEI(x) cvt_f32_to_bf16(x) )==" "\n" |
643 | R"==(#elif DIFF_WEI_DT_U8 )==" "\n" |
644 | R"==(#define TO_DIFF_WEI(x) convert_uchar_sat_rte(x) )==" "\n" |
645 | R"==(#elif DIFF_WEI_DT_S8 )==" "\n" |
646 | R"==(#define TO_DIFF_WEI(x) convert_char_sat_rte(x) )==" "\n" |
647 | R"==(#elif DIFF_WEI_DT_S32 )==" "\n" |
648 | R"==(#define TO_DIFF_WEI(x) convert_int_sat_rte(x) )==" "\n" |
649 | R"==(#else )==" "\n" |
650 | R"==(#define TO_DIFF_WEI(x) (x) )==" "\n" |
651 | R"==(#endif )==" "\n" |
652 | R"==(#endif )==" "\n" |
653 | R"==(#ifdef B_DATA_T )==" "\n" |
654 | R"==(#if B_DT_BF16 )==" "\n" |
655 | R"==(#define B_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
656 | R"==(#define REF_TO_B(x) cvt_f32_to_bf16(x) )==" "\n" |
657 | R"==(#else )==" "\n" |
658 | R"==(#define B_TO_REF(x) (x) )==" "\n" |
659 | R"==(#define REF_TO_B(x) (x) )==" "\n" |
660 | R"==(#endif )==" "\n" |
661 | R"==(#if B_DT_BF16 )==" "\n" |
662 | R"==(#define TO_B(x) cvt_f32_to_bf16(x) )==" "\n" |
663 | R"==(#elif B_DT_U8 )==" "\n" |
664 | R"==(#define TO_B(x) convert_uchar_sat_rte(x) )==" "\n" |
665 | R"==(#elif B_DT_S8 )==" "\n" |
666 | R"==(#define TO_B(x) convert_char_sat_rte(x) )==" "\n" |
667 | R"==(#elif B_DT_S32 )==" "\n" |
668 | R"==(#define TO_B(x) convert_int_sat_rte(x) )==" "\n" |
669 | R"==(#else )==" "\n" |
670 | R"==(#define TO_B(x) (x) )==" "\n" |
671 | R"==(#endif )==" "\n" |
672 | R"==(#endif )==" "\n" |
673 | R"==(#ifdef BIA_DATA_T )==" "\n" |
674 | R"==(#define BIA_DATA2_T CONCAT2(BIA_DATA_T, 2) )==" "\n" |
675 | R"==(#if BIA_DT_BF16 )==" "\n" |
676 | R"==(#define BIA_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
677 | R"==(#define REF_TO_BIA(x) cvt_f32_to_bf16(x) )==" "\n" |
678 | R"==(#else )==" "\n" |
679 | R"==(#define BIA_TO_REF(x) (x) )==" "\n" |
680 | R"==(#define REF_TO_BIA(x) (x) )==" "\n" |
681 | R"==(#endif )==" "\n" |
682 | R"==(#if BIA_DT_BF16 )==" "\n" |
683 | R"==(#define TO_BIA(x) cvt_f32_to_bf16(x) )==" "\n" |
684 | R"==(#elif BIA_DT_U8 )==" "\n" |
685 | R"==(#define TO_BIA(x) convert_uchar_sat_rte(x) )==" "\n" |
686 | R"==(#elif BIA_DT_S8 )==" "\n" |
687 | R"==(#define TO_BIA(x) convert_char_sat_rte(x) )==" "\n" |
688 | R"==(#elif BIA_DT_S32 )==" "\n" |
689 | R"==(#define TO_BIA(x) convert_int_sat_rte(x) )==" "\n" |
690 | R"==(#else )==" "\n" |
691 | R"==(#define TO_BIA(x) (x) )==" "\n" |
692 | R"==(#endif )==" "\n" |
693 | R"==(#endif )==" "\n" |
694 | R"==(#ifdef DST_DATA_T )==" "\n" |
695 | R"==(#define DST_DATA2_T CONCAT2(DST_DATA_T, 2) )==" "\n" |
696 | R"==(#define DST_DATA4_T CONCAT2(DST_DATA_T, 4) )==" "\n" |
697 | R"==(#define DST_DATA8_T CONCAT2(DST_DATA_T, 8) )==" "\n" |
698 | R"==(#define DST_DATA16_T CONCAT2(DST_DATA_T, 16) )==" "\n" |
699 | R"==(#define AS_DST_DATA_T CONCAT2(as_, DST_DATA_T) )==" "\n" |
700 | R"==(#define AS_DST_DATA2_T CONCAT2(as_, DST_DATA2_T) )==" "\n" |
701 | R"==(#define AS_DST_DATA4_T CONCAT2(as_, DST_DATA4_T) )==" "\n" |
702 | R"==(#define AS_DST_DATA8_T CONCAT2(as_, DST_DATA8_T) )==" "\n" |
703 | R"==(#define AS_DST_DATA16_T CONCAT2(as_, DST_DATA16_T) )==" "\n" |
704 | R"==(#if DST_DT_F32 || DST_DT_F16 )==" "\n" |
705 | R"==(#define CONVERT_DST_DATA_T CONCAT2(convert_, DST_DATA_T) )==" "\n" |
706 | R"==(#define CONVERT_DST_DATA2_T CONCAT2(convert_, DST_DATA2_T) )==" "\n" |
707 | R"==(#define CONVERT_DST_DATA4_T CONCAT2(convert_, DST_DATA4_T) )==" "\n" |
708 | R"==(#define CONVERT_DST_DATA8_T CONCAT2(convert_, DST_DATA8_T) )==" "\n" |
709 | R"==(#define CONVERT_DST_DATA16_T CONCAT2(convert_, DST_DATA16_T) )==" "\n" |
710 | R"==(#else )==" "\n" |
711 | R"==(#define CONVERT_DST_DATA_T CONCAT3(convert_, DST_DATA_T, _sat_rte) )==" "\n" |
712 | R"==(#define CONVERT_DST_DATA2_T CONCAT3(convert_, DST_DATA2_T, _sat_rte) )==" "\n" |
713 | R"==(#define CONVERT_DST_DATA4_T CONCAT3(convert_, DST_DATA4_T, _sat_rte) )==" "\n" |
714 | R"==(#define CONVERT_DST_DATA8_T CONCAT3(convert_, DST_DATA8_T, _sat_rte) )==" "\n" |
715 | R"==(#define CONVERT_DST_DATA16_T CONCAT3(convert_, DST_DATA16_T, _sat_rte) )==" "\n" |
716 | R"==(#endif )==" "\n" |
717 | R"==(#if DST_DT_U8 )==" "\n" |
718 | R"==(#define MMAD_DATA_T uint )==" "\n" |
719 | R"==(#define MMAD_DATA4_T uint4 )==" "\n" |
720 | R"==(#define MMAD_DATA8_T uint8 )==" "\n" |
721 | R"==(#elif DST_DT_S8 )==" "\n" |
722 | R"==(#define MMAD_DATA_T int )==" "\n" |
723 | R"==(#define MMAD_DATA4_T int4 )==" "\n" |
724 | R"==(#define MMAD_DATA8_T int8 )==" "\n" |
725 | R"==(#endif )==" "\n" |
726 | R"==(#if DST_DT_U8 || DST_DT_S8 )==" "\n" |
727 | R"==(#define BLOCK_READ_DST2(ptr) \ )==" "\n" |
728 | R"==(AS_DST_DATA2_T(intel_sub_group_block_read_uc2((__global uchar *)ptr)) )==" "\n" |
729 | R"==(#define BLOCK_WRITE_DST2(ptr, v) \ )==" "\n" |
730 | R"==(intel_sub_group_block_write_uc2((__global uchar *)ptr, as_uchar2(v)) )==" "\n" |
731 | R"==(#define BLOCK_READ_DST(ptr) \ )==" "\n" |
732 | R"==(AS_DST_DATA_T(intel_sub_group_block_read_uc((__global uchar *)ptr)) )==" "\n" |
733 | R"==(#define BLOCK_WRITE_DST(ptr, v) \ )==" "\n" |
734 | R"==(intel_sub_group_block_write_uc((__global uchar *)ptr, as_uchar(v)) )==" "\n" |
735 | R"==(#define BLOCK_READ_DST2(ptr) \ )==" "\n" |
736 | R"==(AS_DST_DATA2_T(intel_sub_group_block_read_uc2((__global uchar *)ptr)) )==" "\n" |
737 | R"==(#define BLOCK_WRITE_DST2(ptr, v) \ )==" "\n" |
738 | R"==(intel_sub_group_block_write_uc2((__global uchar *)ptr, as_uchar2(v)) )==" "\n" |
739 | R"==(#define BLOCK_READ_DST4(ptr) \ )==" "\n" |
740 | R"==(AS_DST_DATA4_T(intel_sub_group_block_read_uc4((__global uchar *)ptr)) )==" "\n" |
741 | R"==(#define BLOCK_WRITE_DST4(ptr, v) \ )==" "\n" |
742 | R"==(intel_sub_group_block_write_uc4((__global uchar *)ptr, as_uchar4(v)) )==" "\n" |
743 | R"==(#define BLOCK_READ_DST8(ptr) \ )==" "\n" |
744 | R"==(AS_DST_DATA8_T(intel_sub_group_block_read_uc8((__global uchar *)ptr)) )==" "\n" |
745 | R"==(#define BLOCK_WRITE_DST8(ptr, v) \ )==" "\n" |
746 | R"==(intel_sub_group_block_write_uc8((__global uchar *)ptr, as_uchar8(v)) )==" "\n" |
747 | R"==(#define BLOCK_READ_DST16(ptr) \ )==" "\n" |
748 | R"==(AS_DST_DATA16_T(intel_sub_group_block_read_uc16((__global uchar *)ptr)) )==" "\n" |
749 | R"==(#define BLOCK_WRITE_DST16(ptr, v) \ )==" "\n" |
750 | R"==(intel_sub_group_block_write_uc16((__global uchar *)ptr, as_uchar16(v)) )==" "\n" |
751 | R"==(#elif DST_DT_F16 || DST_DT_BF16 )==" "\n" |
752 | R"==(#define BLOCK_READ_DST(ptr) \ )==" "\n" |
753 | R"==(AS_DST_DATA_T(intel_sub_group_block_read_us((__global ushort *)ptr)) )==" "\n" |
754 | R"==(#define BLOCK_WRITE_DST(ptr, v) \ )==" "\n" |
755 | R"==(intel_sub_group_block_write_us((__global ushort *)ptr, as_ushort(v)) )==" "\n" |
756 | R"==(#define BLOCK_READ_DST2(ptr) \ )==" "\n" |
757 | R"==(AS_DST_DATA2_T(intel_sub_group_block_read_us2((__global ushort *)ptr)) )==" "\n" |
758 | R"==(#define BLOCK_WRITE_DST2(ptr, v) \ )==" "\n" |
759 | R"==(intel_sub_group_block_write_us2((__global ushort *)ptr, as_ushort2(v)) )==" "\n" |
760 | R"==(#define BLOCK_READ_DST4(ptr) \ )==" "\n" |
761 | R"==(AS_DST_DATA4_T(intel_sub_group_block_read_us4((__global ushort *)ptr)) )==" "\n" |
762 | R"==(#define BLOCK_WRITE_DST4(ptr, v) \ )==" "\n" |
763 | R"==(intel_sub_group_block_write_us4((__global ushort *)ptr, as_ushort4(v)) )==" "\n" |
764 | R"==(#define BLOCK_READ_DST8(ptr) \ )==" "\n" |
765 | R"==(AS_DST_DATA8_T(intel_sub_group_block_read_us8((__global ushort *)ptr)) )==" "\n" |
766 | R"==(#define BLOCK_WRITE_DST8(ptr, v) \ )==" "\n" |
767 | R"==(intel_sub_group_block_write_us8((__global ushort *)ptr, as_ushort8(v)) )==" "\n" |
768 | R"==(#define BLOCK_READ_DST16(ptr) \ )==" "\n" |
769 | R"==((DST_DATA16_T)( \ )==" "\n" |
770 | R"==(BLOCK_READ_DST8(ptr), BLOCK_READ_DST8(ptr + 8 * SUB_GROUP_SIZE)) )==" "\n" |
771 | R"==(#define BLOCK_WRITE_DST16(ptr, v) \ )==" "\n" |
772 | R"==(do { \ )==" "\n" |
773 | R"==(BLOCK_WRITE_DST8(ptr, (v).s01234567); \ )==" "\n" |
774 | R"==(BLOCK_WRITE_DST8(ptr + 8 * SUB_GROUP_SIZE, (v).s89abcdef); \ )==" "\n" |
775 | R"==(} while (0) )==" "\n" |
776 | R"==(#elif DST_DT_S32 || DST_DT_F32 )==" "\n" |
777 | R"==(#define BLOCK_READ_DST(ptr) \ )==" "\n" |
778 | R"==(AS_DST_DATA_T(intel_sub_group_block_read((__global uint *)ptr)) )==" "\n" |
779 | R"==(#define BLOCK_WRITE_DST(ptr, v) \ )==" "\n" |
780 | R"==(intel_sub_group_block_write((__global uint *)ptr, as_uint(v)) )==" "\n" |
781 | R"==(#define BLOCK_READ_DST2(ptr) \ )==" "\n" |
782 | R"==(AS_DST_DATA2_T(intel_sub_group_block_read2((__global uint *)ptr)) )==" "\n" |
783 | R"==(#define BLOCK_WRITE_DST2(ptr, v) \ )==" "\n" |
784 | R"==(intel_sub_group_block_write2((__global uint *)ptr, as_uint2(v)) )==" "\n" |
785 | R"==(#define BLOCK_READ_DST4(ptr) \ )==" "\n" |
786 | R"==(AS_DST_DATA4_T(intel_sub_group_block_read4((__global uint *)ptr)) )==" "\n" |
787 | R"==(#define BLOCK_WRITE_DST4(ptr, v) \ )==" "\n" |
788 | R"==(intel_sub_group_block_write4((__global uint *)ptr, as_uint4(v)) )==" "\n" |
789 | R"==(#define BLOCK_READ_DST8(ptr) \ )==" "\n" |
790 | R"==(AS_DST_DATA8_T(intel_sub_group_block_read8((__global uint *)ptr)) )==" "\n" |
791 | R"==(#define BLOCK_WRITE_DST8(ptr, v) \ )==" "\n" |
792 | R"==(intel_sub_group_block_write8((__global uint *)ptr, as_uint8(v)) )==" "\n" |
793 | R"==(#define BLOCK_READ_DST16(ptr) \ )==" "\n" |
794 | R"==((DST_DATA16_T)( \ )==" "\n" |
795 | R"==(BLOCK_READ_DST8(ptr), BLOCK_READ_DST8(ptr + 8 * SUB_GROUP_SIZE)) )==" "\n" |
796 | R"==(#define BLOCK_WRITE_DST16(ptr, v) \ )==" "\n" |
797 | R"==(do { \ )==" "\n" |
798 | R"==(BLOCK_WRITE_DST8(ptr, (v).s01234567); \ )==" "\n" |
799 | R"==(BLOCK_WRITE_DST8(ptr + 8 * SUB_GROUP_SIZE, (v).s89abcdef); \ )==" "\n" |
800 | R"==(} while (0) )==" "\n" |
801 | R"==(#elif DST_DT_F16 || DST_DT_BF16 )==" "\n" |
802 | R"==(#define BLOCK_READ_DST(ptr) \ )==" "\n" |
803 | R"==(AS_DST_DATA_T(intel_sub_group_block_read_us((__global ushort *)ptr)) )==" "\n" |
804 | R"==(#define BLOCK_WRITE_DST(ptr, v) \ )==" "\n" |
805 | R"==(intel_sub_group_block_write_us((__global ushort *)ptr, as_ushort(v)) )==" "\n" |
806 | R"==(#define BLOCK_READ_DST2(ptr) \ )==" "\n" |
807 | R"==(AS_DST_DATA2_T(intel_sub_group_block_read_us2((__global ushort *)ptr)) )==" "\n" |
808 | R"==(#define BLOCK_WRITE_DST2(ptr, v) \ )==" "\n" |
809 | R"==(intel_sub_group_block_write_us2((__global ushort *)ptr, as_short2(v)) )==" "\n" |
810 | R"==(#define BLOCK_READ_DST4(ptr) \ )==" "\n" |
811 | R"==(AS_DST_DATA4_T(intel_sub_group_block_read_us4((__global ushort *)ptr)) )==" "\n" |
812 | R"==(#define BLOCK_WRITE_DST4(ptr, v) \ )==" "\n" |
813 | R"==(intel_sub_group_block_write_us4((__global ushort *)ptr, as_ushort4(v)) )==" "\n" |
814 | R"==(#define BLOCK_READ_DST8(ptr) \ )==" "\n" |
815 | R"==(AS_DST_DATA8_T(intel_sub_group_block_read_us8((__global ushort *)ptr)) )==" "\n" |
816 | R"==(#define BLOCK_WRITE_DST8(ptr, v) \ )==" "\n" |
817 | R"==(intel_sub_group_block_write_us8((__global ushort *)ptr, as_ushort8(v)) )==" "\n" |
818 | R"==(#define BLOCK_READ_DST16(ptr) \ )==" "\n" |
819 | R"==((DST_DATA16_T)( \ )==" "\n" |
820 | R"==(BLOCK_READ_DST8(ptr), BLOCK_READ_DST8(ptr + 8 * SUB_GROUP_SIZE)) )==" "\n" |
821 | R"==(#define BLOCK_WRITE_DST16(ptr, v) \ )==" "\n" |
822 | R"==(do { \ )==" "\n" |
823 | R"==(BLOCK_WRITE_DST8(ptr, (v).s01234567); \ )==" "\n" |
824 | R"==(BLOCK_WRITE_DST8(ptr + 8 * SUB_GROUP_SIZE, (v).s89abcdef); \ )==" "\n" |
825 | R"==(} while (0) )==" "\n" |
826 | R"==(#endif )==" "\n" |
827 | R"==(#if DST_DT_BF16 )==" "\n" |
828 | R"==(#define DST_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
829 | R"==(#define DST_TO_REF2(x) cvt_bf16_to_f32(x) )==" "\n" |
830 | R"==(#define DST_TO_REF8(x) cvt_bf16_to_f32(x) )==" "\n" |
831 | R"==(#define REF_TO_DST(x) cvt_f32_to_bf16(x) )==" "\n" |
832 | R"==(#define REF_TO_DST8(x) cvt_f32_to_bf16(convert_float8(x)) )==" "\n" |
833 | R"==(#elif DST_DT_F16 )==" "\n" |
834 | R"==(#define REF_TO_DST(x) convert_half(x) )==" "\n" |
835 | R"==(#define DST_TO_REF(x) convert_float(x) )==" "\n" |
836 | R"==(#define DST_TO_REF2(x) convert_float2(x) )==" "\n" |
837 | R"==(#define DST_TO_REF8(x) convert_float8(x) )==" "\n" |
838 | R"==(#elif DST_DT_U8 )==" "\n" |
839 | R"==(#define DST_TO_REF(x) (x) )==" "\n" |
840 | R"==(#define DST_TO_REF2(x) (x) )==" "\n" |
841 | R"==(#define DST_TO_REF8(x) (x) )==" "\n" |
842 | R"==(#define REF_TO_DST(x) convert_uchar(x) )==" "\n" |
843 | R"==(#define REF_TO_DST8(x) convert_uchar8(x) )==" "\n" |
844 | R"==(#elif DST_DT_S8 )==" "\n" |
845 | R"==(#define DST_TO_REF(x) (x) )==" "\n" |
846 | R"==(#define DST_TO_REF2(x) (x) )==" "\n" |
847 | R"==(#define DST_TO_REF8(x) (x) )==" "\n" |
848 | R"==(#define REF_TO_DST(x) convert_char(x) )==" "\n" |
849 | R"==(#define REF_TO_DST8(x) convert_char8(x) )==" "\n" |
850 | R"==(#else )==" "\n" |
851 | R"==(#define DST_TO_REF(x) (x) )==" "\n" |
852 | R"==(#define DST_TO_REF2(x) (x) )==" "\n" |
853 | R"==(#define DST_TO_REF8(x) (x) )==" "\n" |
854 | R"==(#define REF_TO_DST(x) (x) )==" "\n" |
855 | R"==(#define REF_TO_DST8(x) (x) )==" "\n" |
856 | R"==(#endif )==" "\n" |
857 | R"==(#if DST_DT_BF16 )==" "\n" |
858 | R"==(#define TO_DST(x) cvt_f32_to_bf16(x) )==" "\n" |
859 | R"==(#define TO_DST2(x) cvt_f32_to_bf16(convert_float2(x)) )==" "\n" |
860 | R"==(#define TO_DST4(x) cvt_f32_to_bf16(convert_float4(x)) )==" "\n" |
861 | R"==(#define TO_DST8(x) cvt_f32_to_bf16(convert_float8(x)) )==" "\n" |
862 | R"==(#elif DST_DT_F16 )==" "\n" |
863 | R"==(#define TO_DST(x) convert_half(x) )==" "\n" |
864 | R"==(#define TO_DST2(x) convert_half2(x) )==" "\n" |
865 | R"==(#define TO_DST4(x) convert_half4(x) )==" "\n" |
866 | R"==(#define TO_DST8(x) convert_half8(x) )==" "\n" |
867 | R"==(#elif DST_DT_U8 )==" "\n" |
868 | R"==(#define TO_DST(x) convert_uchar_sat_rte(x) )==" "\n" |
869 | R"==(#define TO_DST2(x) convert_uchar2_sat_rte(x) )==" "\n" |
870 | R"==(#define TO_DST4(x) convert_uchar4_sat_rte(x) )==" "\n" |
871 | R"==(#define TO_DST8(x) convert_uchar8_sat_rte(x) )==" "\n" |
872 | R"==(#define TO_DST16(x) convert_uchar16_sat_rte(x) )==" "\n" |
873 | R"==(#elif DST_DT_S8 )==" "\n" |
874 | R"==(#define TO_DST(x) convert_char_sat_rte(x) )==" "\n" |
875 | R"==(#define TO_DST2(x) convert_char2_sat_rte(x) )==" "\n" |
876 | R"==(#define TO_DST4(x) convert_char4_sat_rte(x) )==" "\n" |
877 | R"==(#define TO_DST8(x) convert_char8_sat_rte(x) )==" "\n" |
878 | R"==(#define TO_DST16(x) convert_char16_sat_rte(x) )==" "\n" |
879 | R"==(#elif DST_DT_S32 )==" "\n" |
880 | R"==(#define TO_DST(x) convert_int_sat_rte(x) )==" "\n" |
881 | R"==(#define TO_DST2(x) convert_int2_sat_rte(x) )==" "\n" |
882 | R"==(#define TO_DST4(x) convert_int4_sat_rte(x) )==" "\n" |
883 | R"==(#define TO_DST8(x) convert_int8_sat_rte(x) )==" "\n" |
884 | R"==(#elif DST_DT_F32 )==" "\n" |
885 | R"==(#define TO_DST(x) convert_float(x) )==" "\n" |
886 | R"==(#define TO_DST2(x) convert_float2(x) )==" "\n" |
887 | R"==(#define TO_DST4(x) convert_float4(x) )==" "\n" |
888 | R"==(#define TO_DST8(x) convert_float8(x) )==" "\n" |
889 | R"==(#elif DST_DT_F64 )==" "\n" |
890 | R"==(#define TO_DST(x) convert_double(x) )==" "\n" |
891 | R"==(#define TO_DST2(x) convert_double2(x) )==" "\n" |
892 | R"==(#define TO_DST4(x) convert_double4(x) )==" "\n" |
893 | R"==(#define TO_DST8(x) convert_double8(x) )==" "\n" |
894 | R"==(#else )==" "\n" |
895 | R"==(#error "Not expected" )==" "\n" |
896 | R"==(#endif )==" "\n" |
897 | R"==(#endif )==" "\n" |
898 | R"==(#ifdef C_DATA_T )==" "\n" |
899 | R"==(#define C_DATA8_T CONCAT2(C_DATA_T, 8) )==" "\n" |
900 | R"==(#if C_DT_BF16 )==" "\n" |
901 | R"==(#define C_TO_REF(x) cvt_bf16_to_f32(x) )==" "\n" |
902 | R"==(#define C_TO_REF8(x) cvt_bf16_to_f32(x) )==" "\n" |
903 | R"==(#define REF_TO_C(x) cvt_f32_to_bf16(x) )==" "\n" |
904 | R"==(#define REF_TO_C8(x) cvt_f32_to_bf16(convert_float8(x)) )==" "\n" |
905 | R"==(#else )==" "\n" |
906 | R"==(#define C_TO_REF(x) (x) )==" "\n" |
907 | R"==(#define C_TO_REF8(x) (x) )==" "\n" |
908 | R"==(#define REF_TO_C(x) (x) )==" "\n" |
909 | R"==(#define REF_TO_C8(x) (x) )==" "\n" |
910 | R"==(#endif )==" "\n" |
911 | R"==(#if C_DT_BF16 )==" "\n" |
912 | R"==(#define TO_C(x) cvt_f32_to_bf16(x) )==" "\n" |
913 | R"==(#define TO_C8(x) cvt_f32_to_bf16(convert_float8(x)) )==" "\n" |
914 | R"==(#elif C_DT_F16 )==" "\n" |
915 | R"==(#define TO_C(x) convert_half(x) )==" "\n" |
916 | R"==(#define TO_C8(x) convert_half8(x) )==" "\n" |
917 | R"==(#elif C_DT_U8 )==" "\n" |
918 | R"==(#define TO_C(x) convert_uchar_sat_rte(x) )==" "\n" |
919 | R"==(#define TO_C8(x) convert_uchar8_sat_rte(x) )==" "\n" |
920 | R"==(#elif C_DT_S8 )==" "\n" |
921 | R"==(#define TO_C(x) convert_char_sat_rte(x) )==" "\n" |
922 | R"==(#define TO_C8(x) convert_char8_sat_rte(x) )==" "\n" |
923 | R"==(#elif C_DT_S32 )==" "\n" |
924 | R"==(#define TO_C(x) convert_int_sat_rte(x) )==" "\n" |
925 | R"==(#define TO_C8(x) convert_int8_sat_rte(x) )==" "\n" |
926 | R"==(#elif C_DT_F32 )==" "\n" |
927 | R"==(#define TO_C(x) convert_float(x) )==" "\n" |
928 | R"==(#define TO_C8(x) convert_float8(x) )==" "\n" |
929 | R"==(#elif C_DT_F64 )==" "\n" |
930 | R"==(#define TO_C(x) convert_double(x) )==" "\n" |
931 | R"==(#define TO_C8(x) convert_double8(x) )==" "\n" |
932 | R"==(#else )==" "\n" |
933 | R"==(#error "Not expected" )==" "\n" |
934 | R"==(#endif )==" "\n" |
935 | R"==(#endif )==" "\n" |
936 | R"==(#ifdef ACC_DATA_T )==" "\n" |
937 | R"==(#if ACC_DT_F16 )==" "\n" |
938 | R"==(#define TO_ACC(x) convert_half(x) )==" "\n" |
939 | R"==(#elif ACC_DT_F32 )==" "\n" |
940 | R"==(#define TO_ACC(x) convert_float(x) )==" "\n" |
941 | R"==(#elif ACC_DT_F64 )==" "\n" |
942 | R"==(#define TO_ACC(x) convert_double(x) )==" "\n" |
943 | R"==(#elif ACC_DT_S32 )==" "\n" |
944 | R"==(#define TO_ACC(x) convert_int(x) )==" "\n" |
945 | R"==(#else )==" "\n" |
946 | R"==(#error "Unexpected accumulation data type" )==" "\n" |
947 | R"==(#endif )==" "\n" |
948 | R"==(#endif )==" "\n" |
949 | R"==(#ifdef SUM_DATA_T )==" "\n" |
950 | R"==(#define SUM_DATA2_T CONCAT2(SUM_DATA_T, 2) )==" "\n" |
951 | R"==(#define SUM_DATA4_T CONCAT2(SUM_DATA_T, 4) )==" "\n" |
952 | R"==(#define SUM_DATA8_T CONCAT2(SUM_DATA_T, 8) )==" "\n" |
953 | R"==(#define SUM_DATA16_T CONCAT2(SUM_DATA_T, 16) )==" "\n" |
954 | R"==(#define AS_SUM_DATA_T CONCAT2(as_, SUM_DATA_T) )==" "\n" |
955 | R"==(#define AS_SUM_DATA2_T CONCAT2(as_, SUM_DATA2_T) )==" "\n" |
956 | R"==(#define AS_SUM_DATA4_T CONCAT2(as_, SUM_DATA4_T) )==" "\n" |
957 | R"==(#define AS_SUM_DATA8_T CONCAT2(as_, SUM_DATA8_T) )==" "\n" |
958 | R"==(#define AS_SUM_DATA16_T CONCAT2(as_, SUM_DATA16_T) )==" "\n" |
959 | R"==(#if SUM_DT_BF16 )==" "\n" |
960 | R"==(#define SUM_TO_REF cvt_bf16_to_f32 )==" "\n" |
961 | R"==(#else )==" "\n" |
962 | R"==(#define SUM_TO_REF )==" "\n" |
963 | R"==(#endif )==" "\n" |
964 | R"==(#endif )==" "\n" |
965 | R"==(#define OFF_MD_2(prefix, x0, x1, x2, x3, x4, x5) \ )==" "\n" |
966 | R"==(((((x0) / CONCAT2(prefix, _B0_2)) / CONCAT2(prefix, _B0_1) \ )==" "\n" |
967 | R"==(* CONCAT2(prefix, _S0_0)) \ )==" "\n" |
968 | R"==(+ (((x0) / CONCAT2(prefix, _B0_2)) % CONCAT2(prefix, _B0_1) \ )==" "\n" |
969 | R"==(* CONCAT2(prefix, _S0_1)) \ )==" "\n" |
970 | R"==(+ (((x0) % CONCAT2(prefix, _B0_2)) * CONCAT2(prefix, _S0_2)) \ )==" "\n" |
971 | R"==(+ (((x1) / CONCAT2(prefix, _B1_2)) / CONCAT2(prefix, _B1_1) \ )==" "\n" |
972 | R"==(* CONCAT2(prefix, _S1_0)) \ )==" "\n" |
973 | R"==(+ (((x1) / CONCAT2(prefix, _B1_2)) % CONCAT2(prefix, _B1_1) \ )==" "\n" |
974 | R"==(* CONCAT2(prefix, _S1_1)) \ )==" "\n" |
975 | R"==(+ (((x1) % CONCAT2(prefix, _B1_2)) * CONCAT2(prefix, _S1_2)) \ )==" "\n" |
976 | R"==(+ (((x2) / CONCAT2(prefix, _B2_2)) / CONCAT2(prefix, _B2_1) \ )==" "\n" |
977 | R"==(* CONCAT2(prefix, _S2_0)) \ )==" "\n" |
978 | R"==(+ (((x2) / CONCAT2(prefix, _B2_2)) % CONCAT2(prefix, _B2_1) \ )==" "\n" |
979 | R"==(* CONCAT2(prefix, _S2_1)) \ )==" "\n" |
980 | R"==(+ (((x2) % CONCAT2(prefix, _B2_2)) * CONCAT2(prefix, _S2_2)) \ )==" "\n" |
981 | R"==(+ (((x3) / CONCAT2(prefix, _B3_2)) / CONCAT2(prefix, _B3_1) \ )==" "\n" |
982 | R"==(* CONCAT2(prefix, _S3_0)) \ )==" "\n" |
983 | R"==(+ (((x3) / CONCAT2(prefix, _B3_2)) % CONCAT2(prefix, _B3_1) \ )==" "\n" |
984 | R"==(* CONCAT2(prefix, _S3_1)) \ )==" "\n" |
985 | R"==(+ (((x3) % CONCAT2(prefix, _B3_2)) * CONCAT2(prefix, _S3_2)) \ )==" "\n" |
986 | R"==(+ (((x4) / CONCAT2(prefix, _B4_2)) / CONCAT2(prefix, _B4_1) \ )==" "\n" |
987 | R"==(* CONCAT2(prefix, _S4_0)) \ )==" "\n" |
988 | R"==(+ (((x4) / CONCAT2(prefix, _B4_2)) % CONCAT2(prefix, _B4_1) \ )==" "\n" |
989 | R"==(* CONCAT2(prefix, _S4_1)) \ )==" "\n" |
990 | R"==(+ (((x4) % CONCAT2(prefix, _B4_2)) * CONCAT2(prefix, _S4_2)) \ )==" "\n" |
991 | R"==(+ (((x5) / CONCAT2(prefix, _B5_2)) / CONCAT2(prefix, _B5_1) \ )==" "\n" |
992 | R"==(* CONCAT2(prefix, _S5_0)) \ )==" "\n" |
993 | R"==(+ (((x5) / CONCAT2(prefix, _B5_2)) % CONCAT2(prefix, _B5_1) \ )==" "\n" |
994 | R"==(* CONCAT2(prefix, _S5_1)) \ )==" "\n" |
995 | R"==(+ (((x5) % CONCAT2(prefix, _B5_2)) * CONCAT2(prefix, _S5_2))) )==" "\n" |
996 | R"==(#define OFF_MD_3(prefix, x0, x1, x2, x3, x4, x5) \ )==" "\n" |
997 | R"==(((((((x0) / CONCAT2(prefix, _B0_3)) / CONCAT2(prefix, _B0_2)) \ )==" "\n" |
998 | R"==(/ CONCAT2(prefix, _B0_1)) \ )==" "\n" |
999 | R"==(* CONCAT2(prefix, _S0_0)) \ )==" "\n" |
1000 | R"==(+ (((((x0) / CONCAT2(prefix, _B0_3)) / CONCAT2(prefix, _B0_2)) \ )==" "\n" |
1001 | R"==(% CONCAT2(prefix, _B0_1)) \ )==" "\n" |
1002 | R"==(* CONCAT2(prefix, _S0_1)) \ )==" "\n" |
1003 | R"==(+ ((((x0) / CONCAT2(prefix, _B0_3)) % CONCAT2(prefix, _B0_2)) \ )==" "\n" |
1004 | R"==(* CONCAT2(prefix, _S0_2)) \ )==" "\n" |
1005 | R"==(+ (((x0) % CONCAT2(prefix, _B0_3)) * CONCAT2(prefix, _S0_3)) \ )==" "\n" |
1006 | R"==(+ (((((x1) / CONCAT2(prefix, _B1_3)) / CONCAT2(prefix, _B1_2)) \ )==" "\n" |
1007 | R"==(/ CONCAT2(prefix, _B1_1)) \ )==" "\n" |
1008 | R"==(* CONCAT2(prefix, _S1_0)) \ )==" "\n" |
1009 | R"==(+ (((((x1) / CONCAT2(prefix, _B1_3)) / CONCAT2(prefix, _B1_2)) \ )==" "\n" |
1010 | R"==(% CONCAT2(prefix, _B1_1)) \ )==" "\n" |
1011 | R"==(* CONCAT2(prefix, _S1_1)) \ )==" "\n" |
1012 | R"==(+ ((((x1) / CONCAT2(prefix, _B1_3)) % CONCAT2(prefix, _B1_2)) \ )==" "\n" |
1013 | R"==(* CONCAT2(prefix, _S1_2)) \ )==" "\n" |
1014 | R"==(+ (((x1) % CONCAT2(prefix, _B1_3)) * CONCAT2(prefix, _S1_3)) \ )==" "\n" |
1015 | R"==(+ (((((x2) / CONCAT2(prefix, _B2_3)) / CONCAT2(prefix, _B2_2)) \ )==" "\n" |
1016 | R"==(/ CONCAT2(prefix, _B2_1)) \ )==" "\n" |
1017 | R"==(* CONCAT2(prefix, _S2_0)) \ )==" "\n" |
1018 | R"==(+ (((((x2) / CONCAT2(prefix, _B2_3)) / CONCAT2(prefix, _B2_2)) \ )==" "\n" |
1019 | R"==(% CONCAT2(prefix, _B2_1)) \ )==" "\n" |
1020 | R"==(* CONCAT2(prefix, _S2_1)) \ )==" "\n" |
1021 | R"==(+ ((((x2) / CONCAT2(prefix, _B2_3)) % CONCAT2(prefix, _B2_2)) \ )==" "\n" |
1022 | R"==(* CONCAT2(prefix, _S2_2)) \ )==" "\n" |
1023 | R"==(+ (((x2) % CONCAT2(prefix, _B2_3)) * CONCAT2(prefix, _S2_3)) \ )==" "\n" |
1024 | R"==(+ (((((x3) / CONCAT2(prefix, _B3_3)) / CONCAT2(prefix, _B3_2)) \ )==" "\n" |
1025 | R"==(/ CONCAT2(prefix, _B3_1)) \ )==" "\n" |
1026 | R"==(* CONCAT2(prefix, _S3_0)) \ )==" "\n" |
1027 | R"==(+ (((((x3) / CONCAT2(prefix, _B3_3)) / CONCAT2(prefix, _B3_2)) \ )==" "\n" |
1028 | R"==(% CONCAT2(prefix, _B3_1)) \ )==" "\n" |
1029 | R"==(* CONCAT2(prefix, _S3_1)) \ )==" "\n" |
1030 | R"==(+ ((((x3) / CONCAT2(prefix, _B3_3)) % CONCAT2(prefix, _B3_2)) \ )==" "\n" |
1031 | R"==(* CONCAT2(prefix, _S3_2)) \ )==" "\n" |
1032 | R"==(+ (((x3) % CONCAT2(prefix, _B3_3)) * CONCAT2(prefix, _S3_3)) \ )==" "\n" |
1033 | R"==(+ (((((x4) / CONCAT2(prefix, _B4_3)) / CONCAT2(prefix, _B4_2)) \ )==" "\n" |
1034 | R"==(/ CONCAT2(prefix, _B4_1)) \ )==" "\n" |
1035 | R"==(* CONCAT2(prefix, _S4_0)) \ )==" "\n" |
1036 | R"==(+ (((((x4) / CONCAT2(prefix, _B4_3)) / CONCAT2(prefix, _B4_2)) \ )==" "\n" |
1037 | R"==(% CONCAT2(prefix, _B4_1)) \ )==" "\n" |
1038 | R"==(* CONCAT2(prefix, _S4_1)) \ )==" "\n" |
1039 | R"==(+ ((((x4) / CONCAT2(prefix, _B4_3)) % CONCAT2(prefix, _B4_2)) \ )==" "\n" |
1040 | R"==(* CONCAT2(prefix, _S4_2)) \ )==" "\n" |
1041 | R"==(+ (((x4) % CONCAT2(prefix, _B4_3)) * CONCAT2(prefix, _S4_3)) \ )==" "\n" |
1042 | R"==(+ (((((x5) / CONCAT2(prefix, _B5_3)) / CONCAT2(prefix, _B5_2)) \ )==" "\n" |
1043 | R"==(/ CONCAT2(prefix, _B5_1)) \ )==" "\n" |
1044 | R"==(* CONCAT2(prefix, _S5_0)) \ )==" "\n" |
1045 | R"==(+ (((((x5) / CONCAT2(prefix, _B5_3)) / CONCAT2(prefix, _B5_2)) \ )==" "\n" |
1046 | R"==(% CONCAT2(prefix, _B5_1)) \ )==" "\n" |
1047 | R"==(* CONCAT2(prefix, _S5_1)) \ )==" "\n" |
1048 | R"==(+ ((((x5) / CONCAT2(prefix, _B5_3)) % CONCAT2(prefix, _B5_2)) \ )==" "\n" |
1049 | R"==(* CONCAT2(prefix, _S5_2)) \ )==" "\n" |
1050 | R"==(+ (((x5) % CONCAT2(prefix, _B5_3)) * CONCAT2(prefix, _S5_3))) )==" "\n" |
1051 | R"==(#define OFF_MD(prefix, x0, x1, x2, x3, x4, x5) \ )==" "\n" |
1052 | R"==(CONCAT2(OFF_MD_, CONCAT2(prefix, _NLEVELS))(prefix, x0, x1, x2, x3, x4, x5) )==" "\n" |
1053 | R"==(#if SRC_NDIMS == 3 )==" "\n" |
1054 | R"==(#define CONV_SRC_OFF(n, c, d, h, w) OFF_MD(SRC, n, c, w, 0, 0, 0) )==" "\n" |
1055 | R"==(#elif SRC_NDIMS == 4 )==" "\n" |
1056 | R"==(#define CONV_SRC_OFF(n, c, d, h, w) OFF_MD(SRC, n, c, h, w, 0, 0) )==" "\n" |
1057 | R"==(#elif SRC_NDIMS == 5 )==" "\n" |
1058 | R"==(#define CONV_SRC_OFF(n, c, d, h, w) OFF_MD(SRC, n, c, d, h, w, 0) )==" "\n" |
1059 | R"==(#endif )==" "\n" |
1060 | R"==(#if WEI_NDIMS == 3 )==" "\n" |
1061 | R"==(#define CONV_WEI_OFF(g, o, i, d, h, w) OFF_MD(WEI, o, i, w, 0, 0, 0) )==" "\n" |
1062 | R"==(#elif WEI_NDIMS == 4 )==" "\n" |
1063 | R"==(#if WITH_GROUPS == 0 )==" "\n" |
1064 | R"==(#define CONV_WEI_OFF(g, o, i, d, h, w) OFF_MD(WEI, o, i, h, w, 0, 0) )==" "\n" |
1065 | R"==(#else )==" "\n" |
1066 | R"==(#define CONV_WEI_OFF(g, o, i, d, h, w) OFF_MD(WEI, g, o, i, w, 0, 0) )==" "\n" |
1067 | R"==(#endif )==" "\n" |
1068 | R"==(#elif WEI_NDIMS == 5 )==" "\n" |
1069 | R"==(#if WITH_GROUPS == 0 )==" "\n" |
1070 | R"==(#define CONV_WEI_OFF(g, o, i, d, h, w) OFF_MD(WEI, o, i, d, h, w, 0) )==" "\n" |
1071 | R"==(#else )==" "\n" |
1072 | R"==(#define CONV_WEI_OFF(g, o, i, d, h, w) OFF_MD(WEI, g, o, i, h, w, 0) )==" "\n" |
1073 | R"==(#endif )==" "\n" |
1074 | R"==(#elif WEI_NDIMS == 6 )==" "\n" |
1075 | R"==(#define CONV_WEI_OFF(g, o, i, d, h, w) OFF_MD(WEI, g, o, i, d, h, w) )==" "\n" |
1076 | R"==(#endif )==" "\n" |
1077 | R"==(#if DST_NDIMS == 3 )==" "\n" |
1078 | R"==(#define CONV_DST_OFF(n, c, d, h, w) OFF_MD(DST, n, c, w, 0, 0, 0) )==" "\n" |
1079 | R"==(#elif DST_NDIMS == 4 )==" "\n" |
1080 | R"==(#define CONV_DST_OFF(n, c, d, h, w) OFF_MD(DST, n, c, h, w, 0, 0) )==" "\n" |
1081 | R"==(#elif DST_NDIMS == 5 )==" "\n" |
1082 | R"==(#define CONV_DST_OFF(n, c, d, h, w) OFF_MD(DST, n, c, d, h, w, 0) )==" "\n" |
1083 | R"==(#endif )==" "\n" |
1084 | R"==(#if NDIMS == 2 )==" "\n" |
1085 | R"==(#define SRC_OFF(x0, x1, d, h, w) \ )==" "\n" |
1086 | R"==((((x0) % SRC_B0) * SRC_SB0 + ((x0) / SRC_B0) * SRC_S0 \ )==" "\n" |
1087 | R"==(+ ((x1) % SRC_B1) * SRC_SB1 + ((x1) / SRC_B1) * SRC_S1) )==" "\n" |
1088 | R"==(#if WITH_GROUPS == 1 )==" "\n" |
1089 | R"==(#define WEI_OFF(x0, x1, x2, d, h, w) \ )==" "\n" |
1090 | R"==((((x0) % WEI_B0) * WEI_SB0 + ((x0) / WEI_B0) * WEI_S0 \ )==" "\n" |
1091 | R"==(+ ((x1) % WEI_B1) * WEI_SB1 + ((x1) / WEI_B1) * WEI_S1 \ )==" "\n" |
1092 | R"==(+ ((x2) % WEI_B2) * WEI_SB2 + ((x2) / WEI_B2) * WEI_S2) )==" "\n" |
1093 | R"==(#else )==" "\n" |
1094 | R"==(#define WEI_OFF(g, x0, x1, d, h, w) \ )==" "\n" |
1095 | R"==((((x0) % WEI_B0) * WEI_SB0 + ((x0) / WEI_B0) * WEI_S0 \ )==" "\n" |
1096 | R"==(+ ((x1) % WEI_B1) * WEI_SB1 + ((x1) / WEI_B1) * WEI_S1) )==" "\n" |
1097 | R"==(#endif )==" "\n" |
1098 | R"==(#define DST_OFF(x0, x1, d, h, w) \ )==" "\n" |
1099 | R"==((((x0) % DST_B0) * DST_SB0 + ((x0) / DST_B0) * DST_S0 \ )==" "\n" |
1100 | R"==(+ ((x1) % DST_B1) * DST_SB1 + ((x1) / DST_B1) * DST_S1) )==" "\n" |
1101 | R"==(#elif NDIMS == 3 )==" "\n" |
1102 | R"==(#define SRC_OFF(x0, x1, d, h, x2) \ )==" "\n" |
1103 | R"==((((x0) % SRC_B0) * SRC_SB0 + ((x0) / SRC_B0) * SRC_S0 \ )==" "\n" |
1104 | R"==(+ ((x1) % SRC_B1) * SRC_SB1 + ((x1) / SRC_B1) * SRC_S1 \ )==" "\n" |
1105 | R"==(+ ((x2) % SRC_B2) * SRC_SB2 + ((x2) / SRC_B2) * SRC_S2) )==" "\n" |
1106 | R"==(#if WITH_GROUPS == 1 )==" "\n" |
1107 | R"==(#define WEI_OFF(x0, x1, x2, d, h, x3) \ )==" "\n" |
1108 | R"==((((x0) % WEI_B0) * WEI_SB0 + ((x0) / WEI_B0) * WEI_S0 \ )==" "\n" |
1109 | R"==(+ ((x1) % WEI_B1) * WEI_SB1 + ((x1) / WEI_B1) * WEI_S1 \ )==" "\n" |
1110 | R"==(+ ((x2) % WEI_B2) * WEI_SB2 + ((x2) / WEI_B2) * WEI_S2 \ )==" "\n" |
1111 | R"==(+ ((x3) % WEI_B3) * WEI_SB3 + ((x3) / WEI_B3) * WEI_S3) )==" "\n" |
1112 | R"==(#else )==" "\n" |
1113 | R"==(#define WEI_OFF(g, x0, x1, d, h, x2) \ )==" "\n" |
1114 | R"==((((x0) % WEI_B0) * WEI_SB0 + ((x0) / WEI_B0) * WEI_S0 \ )==" "\n" |
1115 | R"==(+ ((x1) % WEI_B1) * WEI_SB1 + ((x1) / WEI_B1) * WEI_S1 \ )==" "\n" |
1116 | R"==(+ ((x2) % WEI_B2) * WEI_SB2 + ((x2) / WEI_B2) * WEI_S2) )==" "\n" |
1117 | R"==(#endif )==" "\n" |
1118 | R"==(#define DST_OFF(x0, x1, d, h, x2) \ )==" "\n" |
1119 | R"==((((x0) % DST_B0) * DST_SB0 + ((x0) / DST_B0) * DST_S0 \ )==" "\n" |
1120 | R"==(+ ((x1) % DST_B1) * DST_SB1 + ((x1) / DST_B1) * DST_S1 \ )==" "\n" |
1121 | R"==(+ ((x2) % DST_B2) * DST_SB2 + ((x2) / DST_B2) * DST_S2) )==" "\n" |
1122 | R"==(#elif NDIMS == 4 )==" "\n" |
1123 | R"==(#define SRC_OFF(x0, x1, d, x2, x3) \ )==" "\n" |
1124 | R"==((((x0) % SRC_B0) * SRC_SB0 + ((x0) / SRC_B0) * SRC_S0 \ )==" "\n" |
1125 | R"==(+ ((x1) % SRC_B1) * SRC_SB1 + ((x1) / SRC_B1) * SRC_S1 \ )==" "\n" |
1126 | R"==(+ ((x2) % SRC_B2) * SRC_SB2 + ((x2) / SRC_B2) * SRC_S2 \ )==" "\n" |
1127 | R"==(+ ((x3) % SRC_B3) * SRC_SB3 + ((x3) / SRC_B3) * SRC_S3) )==" "\n" |
1128 | R"==(#if WITH_GROUPS == 1 )==" "\n" |
1129 | R"==(#define WEI_OFF(x0, x1, x2, d, x3, x4) \ )==" "\n" |
1130 | R"==((((x0) % WEI_B0) * WEI_SB0 + ((x0) / WEI_B0) * WEI_S0 \ )==" "\n" |
1131 | R"==(+ ((x1) % WEI_B1) * WEI_SB1 + ((x1) / WEI_B1) * WEI_S1 \ )==" "\n" |
1132 | R"==(+ ((x2) % WEI_B2) * WEI_SB2 + ((x2) / WEI_B2) * WEI_S2 \ )==" "\n" |
1133 | R"==(+ ((x3) % WEI_B3) * WEI_SB3 + ((x3) / WEI_B3) * WEI_S3 \ )==" "\n" |
1134 | R"==(+ ((x4) % WEI_B4) * WEI_SB4 + ((x4) / WEI_B4) * WEI_S4) )==" "\n" |
1135 | R"==(#else )==" "\n" |
1136 | R"==(#define WEI_OFF(g, x1, x2, d, x3, x4) \ )==" "\n" |
1137 | R"==((((x1) % WEI_B0) * WEI_SB0 + ((x1) / WEI_B0) * WEI_S0 \ )==" "\n" |
1138 | R"==(+ ((x2) % WEI_B1) * WEI_SB1 + ((x2) / WEI_B1) * WEI_S1 \ )==" "\n" |
1139 | R"==(+ ((x3) % WEI_B2) * WEI_SB2 + ((x3) / WEI_B2) * WEI_S2 \ )==" "\n" |
1140 | R"==(+ ((x4) % WEI_B3) * WEI_SB3 + ((x4) / WEI_B3) * WEI_S3) )==" "\n" |
1141 | R"==(#endif )==" "\n" |
1142 | R"==(#define DST_OFF(x0, x1, d, x2, x3) \ )==" "\n" |
1143 | R"==((((x0) % DST_B0) * DST_SB0 + ((x0) / DST_B0) * DST_S0 \ )==" "\n" |
1144 | R"==(+ ((x1) % DST_B1) * DST_SB1 + ((x1) / DST_B1) * DST_S1 \ )==" "\n" |
1145 | R"==(+ ((x2) % DST_B2) * DST_SB2 + ((x2) / DST_B2) * DST_S2 \ )==" "\n" |
1146 | R"==(+ ((x3) % DST_B3) * DST_SB3 + ((x3) / DST_B3) * DST_S3) )==" "\n" |
1147 | R"==(#elif NDIMS == 5 )==" "\n" |
1148 | R"==(#define SRC_OFF(x0, x1, x2, x3, x4) \ )==" "\n" |
1149 | R"==((((x0) % SRC_B0) * SRC_SB0 + ((x0) / SRC_B0) * SRC_S0 \ )==" "\n" |
1150 | R"==(+ ((x1) % SRC_B1) * SRC_SB1 + ((x1) / SRC_B1) * SRC_S1 \ )==" "\n" |
1151 | R"==(+ ((x2) % SRC_B2) * SRC_SB2 + ((x2) / SRC_B2) * SRC_S2 \ )==" "\n" |
1152 | R"==(+ ((x3) % SRC_B3) * SRC_SB3 + ((x3) / SRC_B3) * SRC_S3 \ )==" "\n" |
1153 | R"==(+ ((x4) % SRC_B4) * SRC_SB4 + ((x4) / SRC_B4) * SRC_S4) )==" "\n" |
1154 | R"==(#if WITH_GROUPS == 1 )==" "\n" |
1155 | R"==(#define WEI_OFF(x0, x1, x2, x3, x4, x5) \ )==" "\n" |
1156 | R"==((((x0) % WEI_B0) * WEI_SB0 + ((x0) / WEI_B0) * WEI_S0 \ )==" "\n" |
1157 | R"==(+ ((x1) % WEI_B1) * WEI_SB1 + ((x1) / WEI_B1) * WEI_S1 \ )==" "\n" |
1158 | R"==(+ ((x2) % WEI_B2) * WEI_SB2 + ((x2) / WEI_B2) * WEI_S2 \ )==" "\n" |
1159 | R"==(+ ((x3) % WEI_B3) * WEI_SB3 + ((x3) / WEI_B3) * WEI_S3 \ )==" "\n" |
1160 | R"==(+ ((x4) % WEI_B4) * WEI_SB4 + ((x4) / WEI_B4) * WEI_S4 \ )==" "\n" |
1161 | R"==(+ ((x5) % WEI_B5) * WEI_SB5 + ((x5) / WEI_B5) * WEI_S5) )==" "\n" |
1162 | R"==(#else )==" "\n" |
1163 | R"==(#define WEI_OFF(g, x1, x2, x3, x4, x5) \ )==" "\n" |
1164 | R"==((((x1) % WEI_B0) * WEI_SB0 + ((x1) / WEI_B0) * WEI_S0 \ )==" "\n" |
1165 | R"==(+ ((x2) % WEI_B1) * WEI_SB1 + ((x2) / WEI_B1) * WEI_S1 \ )==" "\n" |
1166 | R"==(+ ((x3) % WEI_B2) * WEI_SB2 + ((x3) / WEI_B2) * WEI_S2 \ )==" "\n" |
1167 | R"==(+ ((x4) % WEI_B3) * WEI_SB3 + ((x4) / WEI_B3) * WEI_S3 \ )==" "\n" |
1168 | R"==(+ ((x5) % WEI_B4) * WEI_SB4 + ((x5) / WEI_B4) * WEI_S4) )==" "\n" |
1169 | R"==(#endif )==" "\n" |
1170 | R"==(#define DST_OFF(x0, x1, x2, x3, x4) \ )==" "\n" |
1171 | R"==((((x0) % DST_B0) * DST_SB0 + ((x0) / DST_B0) * DST_S0 \ )==" "\n" |
1172 | R"==(+ ((x1) % DST_B1) * DST_SB1 + ((x1) / DST_B1) * DST_S1 \ )==" "\n" |
1173 | R"==(+ ((x2) % DST_B2) * DST_SB2 + ((x2) / DST_B2) * DST_S2 \ )==" "\n" |
1174 | R"==(+ ((x3) % DST_B3) * DST_SB3 + ((x3) / DST_B3) * DST_S3 \ )==" "\n" |
1175 | R"==(+ ((x4) % DST_B4) * DST_SB4 + ((x4) / DST_B4) * DST_S4) )==" "\n" |
1176 | R"==(#endif )==" "\n" |
1177 | R"==(#define GWS_OP_ZERO(x, y) 0 )==" "\n" |
1178 | R"==(#define GWS_OP_FIRST(x, y) (x) )==" "\n" |
1179 | R"==(#define GWS_OP_MOD(x, y) ((x) % (y)) )==" "\n" |
1180 | R"==(#define ROUND_UP(a,b) (((a) + (b) - 1) / (b)) )==" "\n" |
1181 | R"==(#define GWS0_GET_ID0() GWS0_OP0((get_global_id(GWS0_IDX0) / GWS0_STRIDE0), ROUND_UP(GWS0_DIM0, GWS0_BLOCK0)) / GWS0_VEC_SIZE0 * GWS0_VEC_SIZE0 * GWS0_BLOCK0 )==" "\n" |
1182 | R"==(#define GWS0_GET_ID1() GWS0_OP1((get_global_id(GWS0_IDX1) / GWS0_STRIDE1), ROUND_UP(GWS0_DIM1, GWS0_BLOCK1)) / GWS0_VEC_SIZE1 * GWS0_VEC_SIZE1 * GWS0_BLOCK1 )==" "\n" |
1183 | R"==(#define GWS0_GET_ID2() GWS0_OP2((get_global_id(GWS0_IDX2) / GWS0_STRIDE2), ROUND_UP(GWS0_DIM2, GWS0_BLOCK2)) / GWS0_VEC_SIZE2 * GWS0_VEC_SIZE2 * GWS0_BLOCK2 )==" "\n" |
1184 | R"==(#define GWS0_GET_ID3() GWS0_OP3((get_global_id(GWS0_IDX3) / GWS0_STRIDE3), ROUND_UP(GWS0_DIM3, GWS0_BLOCK3)) / GWS0_VEC_SIZE3 * GWS0_VEC_SIZE3 * GWS0_BLOCK3 )==" "\n" |
1185 | R"==(#define GWS0_GET_ID4() GWS0_OP4((get_global_id(GWS0_IDX4) / GWS0_STRIDE4), ROUND_UP(GWS0_DIM4, GWS0_BLOCK4)) / GWS0_VEC_SIZE4 * GWS0_VEC_SIZE4 * GWS0_BLOCK4 )==" "\n" |
1186 | R"==(#define GWS0_GET_ID5() GWS0_OP5((get_global_id(GWS0_IDX5) / GWS0_STRIDE5), ROUND_UP(GWS0_DIM5, GWS0_BLOCK5)) / GWS0_VEC_SIZE5 * GWS0_VEC_SIZE5 * GWS0_BLOCK5 )==" "\n" |
1187 | R"==(#define GWS0_GET_BLOCK0() GWS0_BLOCK0 )==" "\n" |
1188 | R"==(#define GWS0_GET_BLOCK1() GWS0_BLOCK1 )==" "\n" |
1189 | R"==(#define GWS0_GET_BLOCK2() GWS0_BLOCK2 )==" "\n" |
1190 | R"==(#define GWS0_GET_BLOCK3() GWS0_BLOCK3 )==" "\n" |
1191 | R"==(#define GWS0_GET_BLOCK4() GWS0_BLOCK4 )==" "\n" |
1192 | R"==(#define GWS0_GET_BLOCK5() GWS0_BLOCK5 )==" "\n" |
1193 | R"==(#define GWS1_GET_ID0() GWS1_OP0((get_global_id(GWS1_IDX0) / GWS1_STRIDE0), ROUND_UP(GWS1_DIM0, GWS1_BLOCK0)) / GWS1_VEC_SIZE0 * GWS1_VEC_SIZE0 * GWS1_BLOCK0 )==" "\n" |
1194 | R"==(#define GWS1_GET_ID1() GWS1_OP1((get_global_id(GWS1_IDX1) / GWS1_STRIDE1), ROUND_UP(GWS1_DIM1, GWS1_BLOCK1)) / GWS1_VEC_SIZE1 * GWS1_VEC_SIZE1 * GWS1_BLOCK1 )==" "\n" |
1195 | R"==(#define GWS1_GET_ID2() GWS1_OP2((get_global_id(GWS1_IDX2) / GWS1_STRIDE2), ROUND_UP(GWS1_DIM2, GWS1_BLOCK2)) / GWS1_VEC_SIZE2 * GWS1_VEC_SIZE2 * GWS1_BLOCK2 )==" "\n" |
1196 | R"==(#define GWS1_GET_ID3() GWS1_OP3((get_global_id(GWS1_IDX3) / GWS1_STRIDE3), ROUND_UP(GWS1_DIM3, GWS1_BLOCK3)) / GWS1_VEC_SIZE3 * GWS1_VEC_SIZE3 * GWS1_BLOCK3 )==" "\n" |
1197 | R"==(#define GWS1_GET_ID4() GWS1_OP4((get_global_id(GWS1_IDX4) / GWS1_STRIDE4), ROUND_UP(GWS1_DIM4, GWS1_BLOCK4)) / GWS1_VEC_SIZE4 * GWS1_VEC_SIZE4 * GWS1_BLOCK4 )==" "\n" |
1198 | R"==(#define GWS1_GET_ID5() GWS1_OP5((get_global_id(GWS1_IDX5) / GWS1_STRIDE5), ROUND_UP(GWS1_DIM5, GWS1_BLOCK5)) / GWS1_VEC_SIZE5 * GWS1_VEC_SIZE5 * GWS1_BLOCK5 )==" "\n" |
1199 | R"==(#define GWS1_GET_BLOCK0() GWS1_BLOCK0 )==" "\n" |
1200 | R"==(#define GWS1_GET_BLOCK1() GWS1_BLOCK1 )==" "\n" |
1201 | R"==(#define GWS1_GET_BLOCK2() GWS1_BLOCK2 )==" "\n" |
1202 | R"==(#define GWS1_GET_BLOCK3() GWS1_BLOCK3 )==" "\n" |
1203 | R"==(#define GWS1_GET_BLOCK4() GWS1_BLOCK4 )==" "\n" |
1204 | R"==(#define GWS1_GET_BLOCK5() GWS1_BLOCK5 )==" "\n" |
1205 | R"==(#define GWS2_GET_ID0() GWS2_OP0((get_global_id(GWS2_IDX0) / GWS2_STRIDE0), ROUND_UP(GWS2_DIM0, GWS2_BLOCK0)) / GWS2_VEC_SIZE0 * GWS2_VEC_SIZE0 * GWS2_BLOCK0 )==" "\n" |
1206 | R"==(#define GWS2_GET_ID1() GWS2_OP1((get_global_id(GWS2_IDX1) / GWS2_STRIDE1), ROUND_UP(GWS2_DIM1, GWS2_BLOCK1)) / GWS2_VEC_SIZE1 * GWS2_VEC_SIZE1 * GWS2_BLOCK1 )==" "\n" |
1207 | R"==(#define GWS2_GET_ID2() GWS2_OP2((get_global_id(GWS2_IDX2) / GWS2_STRIDE2), ROUND_UP(GWS2_DIM2, GWS2_BLOCK2)) / GWS2_VEC_SIZE2 * GWS2_VEC_SIZE2 * GWS2_BLOCK2 )==" "\n" |
1208 | R"==(#define GWS2_GET_ID3() GWS2_OP3((get_global_id(GWS2_IDX3) / GWS2_STRIDE3), ROUND_UP(GWS2_DIM3, GWS2_BLOCK3)) / GWS2_VEC_SIZE3 * GWS2_VEC_SIZE3 * GWS2_BLOCK3 )==" "\n" |
1209 | R"==(#define GWS2_GET_ID4() GWS2_OP4((get_global_id(GWS2_IDX4) / GWS2_STRIDE4), ROUND_UP(GWS2_DIM4, GWS2_BLOCK4)) / GWS2_VEC_SIZE4 * GWS2_VEC_SIZE4 * GWS2_BLOCK4 )==" "\n" |
1210 | R"==(#define GWS2_GET_ID5() GWS2_OP5((get_global_id(GWS2_IDX5) / GWS2_STRIDE5), ROUND_UP(GWS2_DIM5, GWS2_BLOCK5)) / GWS2_VEC_SIZE5 * GWS2_VEC_SIZE5 * GWS2_BLOCK5 )==" "\n" |
1211 | R"==(#define GWS2_GET_BLOCK0() GWS2_BLOCK0 )==" "\n" |
1212 | R"==(#define GWS2_GET_BLOCK1() GWS2_BLOCK1 )==" "\n" |
1213 | R"==(#define GWS2_GET_BLOCK2() GWS2_BLOCK2 )==" "\n" |
1214 | R"==(#define GWS2_GET_BLOCK3() GWS2_BLOCK3 )==" "\n" |
1215 | R"==(#define GWS2_GET_BLOCK4() GWS2_BLOCK4 )==" "\n" |
1216 | R"==(#define GWS2_GET_BLOCK5() GWS2_BLOCK5 )==" "\n" |
1217 | R"==(#define GWS3_GET_ID0() GWS3_OP0((get_global_id(GWS3_IDX0) / GWS3_STRIDE0), ROUND_UP(GWS3_DIM0, GWS3_BLOCK0)) / GWS3_VEC_SIZE0 * GWS3_VEC_SIZE0 * GWS3_BLOCK0 )==" "\n" |
1218 | R"==(#define GWS3_GET_ID1() GWS3_OP1((get_global_id(GWS3_IDX1) / GWS3_STRIDE1), ROUND_UP(GWS3_DIM1, GWS3_BLOCK1)) / GWS3_VEC_SIZE1 * GWS3_VEC_SIZE1 * GWS3_BLOCK1 )==" "\n" |
1219 | R"==(#define GWS3_GET_ID2() GWS3_OP2((get_global_id(GWS3_IDX2) / GWS3_STRIDE2), ROUND_UP(GWS3_DIM2, GWS3_BLOCK2)) / GWS3_VEC_SIZE2 * GWS3_VEC_SIZE2 * GWS3_BLOCK2 )==" "\n" |
1220 | R"==(#define GWS3_GET_ID3() GWS3_OP3((get_global_id(GWS3_IDX3) / GWS3_STRIDE3), ROUND_UP(GWS3_DIM3, GWS3_BLOCK3)) / GWS3_VEC_SIZE3 * GWS3_VEC_SIZE3 * GWS3_BLOCK3 )==" "\n" |
1221 | R"==(#define GWS3_GET_ID4() GWS3_OP4((get_global_id(GWS3_IDX4) / GWS3_STRIDE4), ROUND_UP(GWS3_DIM4, GWS3_BLOCK4)) / GWS3_VEC_SIZE4 * GWS3_VEC_SIZE4 * GWS3_BLOCK4 )==" "\n" |
1222 | R"==(#define GWS3_GET_ID5() GWS3_OP5((get_global_id(GWS3_IDX5) / GWS3_STRIDE5), ROUND_UP(GWS3_DIM5, GWS3_BLOCK5)) / GWS3_VEC_SIZE5 * GWS3_VEC_SIZE5 * GWS3_BLOCK5 )==" "\n" |
1223 | R"==(#define GWS3_GET_BLOCK0() GWS3_BLOCK0 )==" "\n" |
1224 | R"==(#define GWS3_GET_BLOCK1() GWS3_BLOCK1 )==" "\n" |
1225 | R"==(#define GWS3_GET_BLOCK2() GWS3_BLOCK2 )==" "\n" |
1226 | R"==(#define GWS3_GET_BLOCK3() GWS3_BLOCK3 )==" "\n" |
1227 | R"==(#define GWS3_GET_BLOCK4() GWS3_BLOCK4 )==" "\n" |
1228 | R"==(#define GWS3_GET_BLOCK5() GWS3_BLOCK5 )==" "\n" |
1229 | R"==(#define KERNEL_ATTR_SG0 \ )==" "\n" |
1230 | R"==(__attribute__((reqd_work_group_size( \ )==" "\n" |
1231 | R"==(GWS_LWS0_DEFAULT, GWS_LWS1_DEFAULT, GWS_LWS2_DEFAULT))) )==" "\n" |
1232 | R"==(#define KERNEL_ATTR_SG1 \ )==" "\n" |
1233 | R"==(KERNEL_ATTR_SG0 \ )==" "\n" |
1234 | R"==(__attribute__((intel_reqd_sub_group_size(GWS_SGS_DEFAULT))) )==" "\n" |
1235 | R"==(#define KERNEL_ATTR CONCAT2(KERNEL_ATTR_SG, GWS_WITH_SG_DEFAULT) )==" "\n" |
1236 | R"==(#define NAMED_KERNEL_ATTR_SG0(name) \ )==" "\n" |
1237 | R"==(__attribute__((reqd_work_group_size(CONCAT2(GWS_LWS0_, name), \ )==" "\n" |
1238 | R"==(CONCAT2(GWS_LWS1_, name), CONCAT2(GWS_LWS2_, name)))) )==" "\n" |
1239 | R"==(#define NAMED_KERNEL_ATTR_SG1(name) \ )==" "\n" |
1240 | R"==(NAMED_KERNEL_ATTR_SG0(name) \ )==" "\n" |
1241 | R"==(__attribute__((intel_reqd_sub_group_size(CONCAT2(GWS_SGS_, name)))) )==" "\n" |
1242 | R"==(#define NAMED_KERNEL_ATTR(name) \ )==" "\n" |
1243 | R"==(CONCAT2(NAMED_KERNEL_ATTR_SG, CONCAT2(GWS_WITH_SG_, name))(name) )==" "\n" |
1244 | R"==(#define MAYBE_SKIP_NON_UNIFORM_WG() \ )==" "\n" |
1245 | R"==(do { \ )==" "\n" |
1246 | R"==(if ((GWS_0 != GWS_ORIG_0) && (GWS_ORIG_0 % LWS_0 != 0) \ )==" "\n" |
1247 | R"==(&& (get_global_id(0) >= GWS_ORIG_0)) \ )==" "\n" |
1248 | R"==(return; \ )==" "\n" |
1249 | R"==(if ((GWS_1 != GWS_ORIG_1) && (GWS_ORIG_1 % LWS_1 != 0) \ )==" "\n" |
1250 | R"==(&& (get_global_id(1) >= GWS_ORIG_1)) \ )==" "\n" |
1251 | R"==(return; \ )==" "\n" |
1252 | R"==(if ((GWS_2 != GWS_ORIG_2) && (GWS_ORIG_2 % LWS_2 != 0) \ )==" "\n" |
1253 | R"==(&& (get_global_id(2) >= GWS_ORIG_2)) \ )==" "\n" |
1254 | R"==(return; \ )==" "\n" |
1255 | R"==(} while (0) )==" "\n" |
1256 | R"==(#endif )==" "\n" |
1257 | R"==(#if SRC_DT_U8 == 1 )==" "\n" |
1258 | R"==(#define SRC_DT_ALIAS UCHAR )==" "\n" |
1259 | R"==(#elif SRC_DT_S8 == 1 )==" "\n" |
1260 | R"==(#define SRC_DT_ALIAS CHAR )==" "\n" |
1261 | R"==(#elif SRC_DT_F16 == 1 )==" "\n" |
1262 | R"==(#define SRC_DT_ALIAS HALF )==" "\n" |
1263 | R"==(#elif SRC_DT_BF16 == 1 )==" "\n" |
1264 | R"==(#define SRC_DT_ALIAS BFLOAT )==" "\n" |
1265 | R"==(#elif SRC_DT_F32 == 1 )==" "\n" |
1266 | R"==(#define SRC_DT_ALIAS FLOAT )==" "\n" |
1267 | R"==(#endif )==" "\n" |
1268 | R"==(#if DST_DT_U8 == 1 )==" "\n" |
1269 | R"==(#define DST_DT_ALIAS UCHAR )==" "\n" |
1270 | R"==(#elif DST_DT_S8 == 1 )==" "\n" |
1271 | R"==(#define DST_DT_ALIAS CHAR )==" "\n" |
1272 | R"==(#elif DST_DT_F16 == 1 )==" "\n" |
1273 | R"==(#define DST_DT_ALIAS HALF )==" "\n" |
1274 | R"==(#elif DST_DT_BF16 == 1 )==" "\n" |
1275 | R"==(#define DST_DT_ALIAS BFLOAT )==" "\n" |
1276 | R"==(#elif DST_DT_F32 == 1 )==" "\n" |
1277 | R"==(#define DST_DT_ALIAS FLOAT )==" "\n" |
1278 | R"==(#endif )==" "\n" |
1279 | R"==(#define ALIAS(prefix) CONCAT2(prefix, _DT_ALIAS) )==" "\n" |
1280 | R"==(#define BLOCK1_T uchar )==" "\n" |
1281 | R"==(#define BLOCK2_T ushort )==" "\n" |
1282 | R"==(#define BLOCK4_T uint )==" "\n" |
1283 | R"==(#define BLOCK8_T ulong )==" "\n" |
1284 | R"==(#define BLOCK_T(alias) CONCAT3(BLOCK, SIZEOF(alias), _T) )==" "\n" |
1285 | R"==(#define BLOCK1_ALIAS UCHAR )==" "\n" |
1286 | R"==(#define BLOCK2_ALIAS USHORT )==" "\n" |
1287 | R"==(#define BLOCK4_ALIAS UINT )==" "\n" |
1288 | R"==(#define BLOCK_ALIAS(prefix) CONCAT3(BLOCK, SIZEOF(ALIAS(prefix)), _ALIAS) )==" "\n" |
1289 | R"==(#define SIZEOF_UCHAR 1 )==" "\n" |
1290 | R"==(#define SIZEOF_CHAR 1 )==" "\n" |
1291 | R"==(#define SIZEOF_BFLOAT 2 )==" "\n" |
1292 | R"==(#define SIZEOF_HALF 2 )==" "\n" |
1293 | R"==(#define SIZEOF_FLOAT 4 )==" "\n" |
1294 | R"==(#define SIZEOF(alias) CONCAT2(SIZEOF_, alias) )==" "\n" |
1295 | R"==(#define READ_UCHAR8 intel_sub_group_block_read_uc8 )==" "\n" |
1296 | R"==(#define READ_USHORT8 intel_sub_group_block_read_us8 )==" "\n" |
1297 | R"==(#define READ_UINT8 intel_sub_group_block_read8 )==" "\n" |
1298 | R"==(#define READ_BLOCK8(prefix, ptr) READ_BLOCK_N(prefix, 8)(ptr) )==" "\n" |
1299 | R"==(#define READ_BLOCK_N(prefix, n) CONCAT3(READ_, BLOCK_ALIAS(prefix), n) )==" "\n" |
1300 | R"==(#define WRITE_UCHAR8 intel_sub_group_block_write_uc8 )==" "\n" |
1301 | R"==(#define WRITE_USHORT8 intel_sub_group_block_write_us8 )==" "\n" |
1302 | R"==(#define WRITE_UINT8 intel_sub_group_block_write8 )==" "\n" |
1303 | R"==(#define WRITE_BLOCK8(prefix, ptr, val) WRITE_BLOCK_N(prefix, 8)(ptr, val) )==" "\n" |
1304 | R"==(#define WRITE_BLOCK_N(prefix, n) CONCAT3(WRITE_, BLOCK_ALIAS(prefix), n) )==" "\n" |
1305 | R"==(#define AS_UCHAR8 as_uchar8 )==" "\n" |
1306 | R"==(#define AS_CHAR8 as_char8 )==" "\n" |
1307 | R"==(#define AS_HALF8 as_half8 )==" "\n" |
1308 | R"==(#define AS_USHORT8 as_ushort8 )==" "\n" |
1309 | R"==(#define AS_BFLOAT8 as_ushort8 )==" "\n" |
1310 | R"==(#define AS_FLOAT8 as_float8 )==" "\n" |
1311 | R"==(#define AS_INT8 as_int8 )==" "\n" |
1312 | R"==(#define AS_UINT8 as_uint8 )==" "\n" |
1313 | R"==(#define BLOCK_TO_DATA8(prefix, val) BLOCK_TO_DATA_N(prefix, 8)(val) )==" "\n" |
1314 | R"==(#define BLOCK_TO_DATA_N(prefix, n) CONCAT3(AS_, ALIAS(prefix), n) )==" "\n" |
1315 | R"==(#define DATA_TO_BLOCK8(prefix, val) DATA_TO_BLOCK_N(prefix, 8)(val) )==" "\n" |
1316 | R"==(#define DATA_TO_BLOCK_N(prefix, n) CONCAT3(AS_, BLOCK_ALIAS(prefix), n) )==" "\n" |
1317 | R"==(#define UCHAR_TO_FLOAT1 convert_float )==" "\n" |
1318 | R"==(#define UCHAR_TO_FLOAT8 convert_float8 )==" "\n" |
1319 | R"==(#define FLOAT_TO_UCHAR1 convert_uchar_sat_rte )==" "\n" |
1320 | R"==(#define FLOAT_TO_UCHAR8 convert_uchar8_sat_rte )==" "\n" |
1321 | R"==(#define CHAR_TO_FLOAT1 convert_float )==" "\n" |
1322 | R"==(#define CHAR_TO_FLOAT8 convert_float8 )==" "\n" |
1323 | R"==(#define FLOAT_TO_CHAR1 convert_char_sat_rte )==" "\n" |
1324 | R"==(#define FLOAT_TO_CHAR8 convert_char8_sat_rte )==" "\n" |
1325 | R"==(#define HALF_TO_FLOAT1 convert_float )==" "\n" |
1326 | R"==(#define HALF_TO_FLOAT8 convert_float8 )==" "\n" |
1327 | R"==(#define FLOAT_TO_HALF1 convert_half )==" "\n" |
1328 | R"==(#define FLOAT_TO_HALF8 convert_half8 )==" "\n" |
1329 | R"==(#define BFLOAT_TO_FLOAT1 cvt_bf16_to_f32 )==" "\n" |
1330 | R"==(#define BFLOAT_TO_FLOAT8 cvt_bf16_to_f32 )==" "\n" |
1331 | R"==(#define FLOAT_TO_BFLOAT1 cvt_f32_to_bf16 )==" "\n" |
1332 | R"==(#define FLOAT_TO_BFLOAT8 cvt_f32_to_bf16 )==" "\n" |
1333 | R"==(#define FLOAT_TO_FLOAT1 convert_float )==" "\n" |
1334 | R"==(#define FLOAT_TO_FLOAT8 convert_float8 )==" "\n" |
1335 | R"==(#define DATA_TO_FLOAT(prefix, val) DATA_TO_FLOAT_N(prefix, 1)(val) )==" "\n" |
1336 | R"==(#define DATA_TO_FLOAT8(prefix, val) DATA_TO_FLOAT_N(prefix, 8)(val) )==" "\n" |
1337 | R"==(#define DATA_TO_FLOAT_N(prefix, n) CONCAT3(ALIAS(prefix), _TO_FLOAT, n) )==" "\n" |
1338 | R"==(#define FLOAT_TO_DATA(prefix, val) FLOAT_TO_DATA_N(prefix, 1)(val) )==" "\n" |
1339 | R"==(#define FLOAT_TO_DATA8(prefix, val) FLOAT_TO_DATA_N(prefix, 8)(val) )==" "\n" |
1340 | R"==(#define FLOAT_TO_DATA_N(prefix, n) CONCAT3(FLOAT_TO_, ALIAS(prefix), n) )==" "\n" |
1341 | R"==()==" ; |
1342 | } |
1343 | } |
1344 | } |
1345 | } |